The Design of SMS Based Heterogeneous Mobile Botnet

合集下载

An Integrated Framework for Optimization Under Uncertainty Using Inverse Reliability Strategy

Xiaoping Du Department of Mechanical and AerospaceEngineering,University of Missouri–Rolla,Rolla,MO65409e-mail:dux@Agus SudjiantoFord Motor Company,Dearborn,MI48121-4091e-mail:asudjian@Wei Chen* Department of Mechanical Engineering,Northwestern University,Evanston,IL60208-3111 e-mail:weichen@ An Integrated Framework for Optimization Under Uncertainty Using Inverse Reliability Strategy In this work,we propose an integrated framework for optimization under uncertainty that can bring both the design objective robustness and the probabilistic design constraints into account.The fundamental development of this work is the employment of an inverse reliability strategy that uses percentile performance for assessing both the objective ro-bustness and probabilistic constraints.The percentile formulation for objective robustness provides us an accurate evaluation of the variation of an objective performance and a probabilistic measurement of the robustness.We can obtain more reasonable compound noise combinations for a robust design objective compared to using the traditional ap-proach proposed by Taguchi.The proposed formulation is very efﬁcient to solve since it only needs to evaluate the constraint functions at the required reliability levels.The other major development of this work is a new search algorithm for the Most Probable Point of Inverse Reliability(MPPIR)that can be used to efﬁciently evaluate percentile perfor-mances for both robustness and reliability assessments.Multiple strategies are employed in the MPPIR search,including using the steepest ascent direction and an arc search.The algorithm is applicable to general non-concave and non-convex performance functions of random variables following any continuous distributions.The effectiveness of the MPPIR search algorithm is veriﬁed using example problems.Overall,an engineering example on integrated robust and reliability design of a vehicle combustion engine piston is used to illustrate the beneﬁts of our proposed method.͓DOI:10.1115/1.1759358͔1IntroductionRecent years have seen many developments of methods for design under uncertainty.Among these developments,Robust de-sign͓1–2͔and reliability-based design͓3͔represent two major paradigms for design under uncertainty.It should be pointed out that the emphases of these two paradigms are different.Robust design is a method for improving the quality of a product through minimizing the effect of variation without eliminating the causes ͓4͔.It emphasizes on achieving the robustness of performance͑for design objective͒.On the other hand,the reliability-based designapproach focuses on maintaining design feasibility͑for designconstraints͒at expected probabilistic levels.It is our belief that fordesign under uncertainty,the needs of robustness and reliabilityshould be integrated.A common challenge that designers face when using either ro-bust design or reliability-based design is the computational ex-pense.Although methods have been developed for improving thecomputational efﬁciency of either robust design or reliability-based design,these methods are distinctly different because of thedifferent emphases these two paradigms have.Speciﬁcally,underthe robust design paradigm,mean and variance of performanceare evaluated for assessing the design objective and the probabi-listic constraints are simpliﬁed using either the worst-case sce-nario͑sensitivity analysis based͒or the moment-matching formu-lation͓4–10͔.The commonly used method to evaluate theperformance deviation͑or variance͒in robust design is theﬁrstorder Taylor expansion.If the variances of random variables arelarge and the performance function is highly nonlinear,this ap-proach may result in large errors.The use of Monte Carlo simu-lation for evaluating probabilistic characteristics is generally notaffordable in many design applications.Under the reliability-based design paradigm,methods have been developed for efﬁciently assessing the probability of con-straints being feasible͑or called reliability͒.Many of these meth-ods are based on the concept of the Most Probable Point͓11–15͔, which emphasizes on assessing the tail performance of a probabi-listic constraint.In reliability-based design,deterministic objec-tives such as the performance at the mean values of random vari-ables are often used.To overcome the difﬁculties and inefﬁciency associated with nested double-loop procedures,sequential single-loop methods that separate the inner probabilistic assessment loop and the outer optimization loop have been proposed͓16–19͔.It is our belief that both robustness and reliability are desired characteristics for design under uncertainty.Therefore these two paradigms need to be integrated in a uniﬁed probabilistic optimi-zation formulation.Although attempts have been made to inte-grate the robustness into the reliability-based design͓20,21͔,none of the existing works addresses the development of efﬁcient com-putational techniques to facilitate the assessments of both robust-ness and reliability characteristics in searching the probabilistic optimal solution.In this work,we propose an integrated frame-work for optimization under uncertainty that can efﬁciently bring both the design objective robustness and the probabilistic design constraints into account.Two major developments are involved. The fundamental development is the employment of an inverse reliability strategy͓17,18,22–24͔that uses percentile performance for assessing both the robustness objective and probabilistic con-straints.The probabilistic constraints are formulated as inverse reliability constraints,which are assessed by equivalent percentile performances͑inverse reliability formulation͒;while the robust-ness is achieved through a design objective in which the variation of a design performance is approximately evaluated through the percentile performance difference between the right and left tails of performance distribution.Corresponding to the use of percen-tile performance,the other major development is a new search algorithm of Most Probable Point of Inverse Reliability͑MPPIR͒. The new algorithm is used to efﬁciently evaluate the robustness and reliability in the proposed formulation.In the remaining part*Corresponding author.Contributed by the Design Automation Committee for publication in the J OUR-NAL OF M ECHANICAL D ESIGN.Manuscript received January2003;revised January2004.Associate Editor:G.M.Fadel.562ÕVol.126,JULY2004Copyright©2004by ASME Transactions of the ASMEof this paper,we will demonstrate the beneﬁts of the proposed integrated framework for optimization under uncertainty and the effectiveness of the MPPIR search algorithm.2A General Design Model Under ConcertaintyA typical design model under uncertainty is given by:Minimize:f͑v gob j͒Design Variable DVϭ͕d,␯x͖(1) Subject to:Prob͕g i͑d,X,P͒р0͖у␣i,iϭ1,2,...,m, In the above model,f is the design objective,which is a func-tion of the probabilistic characteristic v gob j of the objective per-formance g ob j(d,X,P).The probabilistic characteristic v gob j couldinclude the mean,the standard deviation of g ob j,or the combina-tion of both.The probabilistic design objective is to minimize f.dis the vector of deterministic design variables or deterministiccontrol factors.X is the vector of random design variables orrandom control factors.P is the vector of random design param-eters or noise factors.The difference between a design variable ͑either deterministic or random͒and a design parameter is that the former is changeable and controllable by a designer in a designprocess while the latter is not.The decision variables in Eq.͑1͒are d and the distribution parameters␯x of random design vari-ables X.Examples of the distribution parameters␯x include the mean␮x,the standard deviation␴x,etc.g i(d,X,P)(i ϭ1,2,...,m)are constraint functions;Prob͕•͖denotes a prob-ability while␣i(iϭ1,2,...,m)stand for desired probabilities of constraint satisfaction;m is the number of constraints.Note that both the objective performance g ob j and constraint performance g i are performance variables.In the remainder of this paper,we use g to denote any performance variables.In the above design model,the design feasibility is formulatedas the probability of constraint satisfaction g(d,X,P)р0largerthan or equal to a desired probability␣.Usually this probability Prob͕g i(d,X,P)р0͖is called reliability.As shown in Fig.1,the probability of g(d,X,P)р0is the area underneath the curve of probability density function͑PDF͒of g for gр0,and this area should be greater than or equal to␣.In a robust design,the robustness of a design objective can beachieved by simultaneously‘‘optimizing the mean performance ␮ob j’’and‘‘minimizing the performance variance␴ob j’’͓25͔.The performance g ob j(d,X,P)is a function of all random variables.Its mean value␮ob j and variance␴ob j2are to be minimized.The form of the objective can be expressed asmin f͑␮ob j,␴ob j͒.(2) Different from robust design,the emphasis of the reliability-based design is on maintaining the reliability of a constraint͑de-sign feasibility requirement͒.Usually only nominal values are considered for the objective,which is calculated at the means of random variables.Therefore in reliability-based design,the objec-tive is often represented by the nominal value of g ob j,i.e.,min g ob j͑d,␮X,␮P͒(3) In this work,a uniﬁed probabilistic optimization formulation is used to integrate the robustness and reliability considerations.As shown in the following model,the robustness requirement is cap-tured by the design objective while the reliability considerations are modeled with probabilistic constraints.min f͑␮ob j,␴ob j͒Design Variable DVϭ͕d,␯x͖(4) Subject to:Prob͕g i͑d,X,P͒р0͖у␣i,iϭ1,2,...,m, Equation͑4͒represents a multicriteria optimization problem where the tradeoff needs to be made between optimizing the mean performance and minimizing the performance variation͓25͔.How to construct an objective function representing designer’s prefer-ence in making the tradeoff is not the focus of this study.Here we assume that a single objective function is constructed based on both the mean and variance criteria.In this work,an inverse reli-ability strategy is proposed to reformulate the above optimization model,so that the robustness and reliability assessments can be treated in a uniﬁed manner.3An Inverse Reliability Strategy for Reformulating the Optimization Model Under UncertaintyAn inverse reliability strategy is proposed in this work to refor-mulate the probabilistic optimization formulation shown in Eq.͑4͒.This development is motivated by the need for developingcomputationally efﬁcient techniques for solving the integrated probabilistic optimization model and the need for providing a more accurate assessment of performance dispersion in improving system robustness.In conventional reliability analysis,given a prespeciﬁed performance,which is called limit state͓26͔in the ﬁeld of structural reliability,one is interested inﬁnding the prob-ability(reliability)of the performance greater or less than that prespeciﬁed performance.With inverse reliability or called per-centile formulation,we will focus onﬁnding a speciﬁc perfor-mance that corresponds to a given reliability.This task is consid-ered as solving an inverse reliability problem͓18,19͔.In this work,we employ inverse reliability formulations for assessing both the robustness objective and the probabilistic constraints. 3.1Modeling Design Feasibility by Inverse Reliability Strategy.Du and Chen͓6͔discussed commonly used tech-niques for modeling design feasibility under uncertainty and they concluded that the ideal technique is the probabilistic formulation presented in Eq.͑4͒.However,to use Eq.͑4͒,we need to evaluate the reliability Prob͕g i(d,X,P)р0͖for each probabilistic function g i(d,X,P).In presence of multiple constraints,some constraints may never be active and consequently their reliabilities are ex-tremely high͑approaching1.0͒.Although these constraints are the least critical,the evaluations of these reliabilities will unfortu-nately dominate the computational effort in probabilistic optimi-zation.The solution to this problem is to perform the reliability assessment only up to the necessary level͓24͔.Hence,a formula-tion of percentile performance͑inverse reliability͒has been pro-posed to replace the reliability formulation͓17–19,24͔.The per-centile performance formulation is shown as:g␣р0,(5) where g␣is the␣-percentile performance of g(d,X,P),namely,Prob͕g͑d,X,P͒рg␣͖ϭ␣(6) Equation͑5͒indicates that the probability of g(d,X,P)less than or equal to the␣-percentile performance g␣is exactly equal to the desired reliability␣.The concept is demonstrated in Fig.2. If the shaded area,the probability at the left side of Eq.͑6͒,is Fig.1The concept of reliabilityJournal of Mechanical Design JULY2004,Vol.126Õ563equal to the desired reliability ␣,then the point g ␣on g axis is called the ␣-percentile value of function g .From Fig.2we see that,g ␣р0indicates that Prob ͕g i (d ,X ,P )р0͖у␣,which means that the probabilistic constraint is feasible.With the transforma-tion to inverse reliability,the original constraints that require re-liability assessments are now converted to equivalent constraints that evaluate the ␣-percentile performance.Instead of checking the actual reliability,the location of g ␣will now determine the feasibility of a constraint.For simplicity,we use percentile perfor-mance to stand for ␣-percentile performance.It has also been shown that with the percentile formulation we can avoid singular-ity problems which may occur in solving a direct reliability model ͑Eq.͑4͒͒during the iterative reliability assessment procedure ͓24͔.3.2Modeling Design Objective Robustness by Inverse Re-liability Strategy.A robust design objective reﬂects the need for shrinking the dispersion of a performance.As mentioned in Section 1,the use of Taylor expansion is not very accurate in estimating the standard deviation of a performance.With the original Taguchi’s robust design method,‘‘compound noise’’is used to assess high or low quality performances based on combi-nations of several noise factors ͓2͔.The robust design objective is then represented by minimizing the difference of performances at two compound noise combinations.The strategy of compound noise could signiﬁcantly reduce the number of experiments ͑or number of simulations ͒since the performance is evaluated only at two points which supposedly correspond to the highest and lowest quality.However,certain conditions need to be satisﬁed to use this ad hoc approach ͓2͔.For example,͑1͒we must know what the major noise factors are;͑2͒we must know the directionality ͑posi-tive or negative ͒of their effects on the performance;and ͑3͒the directionality of those effects of noise factors should not depend on the settings of the control factors.If the last two conditions are violated,the effect of one noise factor may get compensated for by another noise factor and then the robust design based on the compound noise can give confusing and misleading results.Tagu-chi suggested a typical value of Ϯͱ3/2␴for noise levels,which does not always generate the highest and lowest quality perfor-mance and can lead to wrong conclusion in design selection ͓2͔.To utilize the idea of compound noise but to overcome the aforementioned drawbacks,we propose to use percentile perfor-mance difference to represent the variation of a performance.The percentile performance difference is given by⌬g ob j ␣1␣2ϭg ob j ␣2Ϫg ob j␣1,(7)in which ␣1and ␣2are reliability levels or the cumulative distri-bution functions ͑CDFs ͒of g given byProb ͕f ͑d ,X ,P ͒рg ob j ␣i͖ϭ␣i͑i ϭ1,2͒(8)␣1is a left-tail CDF,for example,0.05or 0.01,which repre-sents the performance at the left tail of its distribution and ␣2is a right-tail CDF,for example,0.95or 0.99.Percentile performances g ob j ␣1(d ,X ,P )and g ob j␣2(d ,X ,P )represent high and low ͑or low and high ͒system qualities respectively.From Fig.3we see that the percentile performance difference is the distance between ␣1per-centile performance and ␣2percentile performance.As shown in Fig.3,minimizing the percentile performance difference helps to shrink the range of the distribution.It should be noted that when a performance distribution is not unimodal,minimizing the percen-tile performance difference at two tails may not decrease the vari-ance of performance distribution.Our method will not be appli-cable for such cases which are often rare in design applications.Different from Taguchi’s approach of compound noise factor,where only random parameters ͑noise factors ͒are involved,our approach of compound noise factor includes a combination of all random variables,namely,random control factors X and noise factors P ͑therefore we call a compound noise factor ‘‘compound noise setting’’in the subsequent sections ͒.There are several advantages of using percentile performance difference to replace the conventional performance variance ͑or standard deviation ͒for robustness assessment.One major advan-tage is that a percentile performance is related to the probability at the tail areas of a performance distribution and therefore it carries more information than the standard deviation such as it could indicate the skewness of a distribution ͑see the example in Section 5͒,while the standard deviation only captures the dispersion around the mean value.Also with percentile formulation,we can immediately know to what extent or at what conﬁdence level the design robustness is achieved.This conﬁdence level is given by ␣2Ϫ␣1.The other major advantage is related to the computa-tional efﬁciency achieved by using inverse reliability assessments ͑percentile evaluations ͒for both robustness objective and proba-bilistic constraints ͓19͔,with more details in Section bined with the concept of the Most Probable Point ͑MPP ͒,the percentile formulation gives us reasonable compound noise settings in ro-bustness evaluations.In summary,using the inverse reliability strategy,the uniﬁed probabilistic optimization model for integrated robustness and re-liability design becomesmin ͕␮g ob j ,⌬g ob j ␣1␣2͖Design Variable DV ϭ͕d ,␯x ͖(9)Subject to:g i ␣͑d ,X ,P ͒р0,i ϭ1,2,...,m .4The Inverse Reliability Assessment MethodTo solve Eq.͑9͒efﬁciently,an efﬁcient search algorithm for the Most Probable Point of Inverse Reliability ͑MPPIR ͒is developed in this work to evaluate the percentile performances.We will ﬁrst discuss our proposed MPPIR search algorithm and then explain how the MPPIR is related to the compound noise setting.4.1Background of MPP Search for the Inverse Reliability Problem.The MPP concept was originally developed in the structural reliability area ͓3͔with the purpose of reliability assess-ment.With the MPP approach,the random variables Y ϭ(X ,P)Fig.2An ␣-percentile of a constraintfunctionFig.3Percentile difference for shrinking the distribution564ÕVol.126,JULY 2004Transactions of the ASMEare transformed into an independent and standardized normal space Uϭ(U X,U P).The transformation is given by͓27͔,U iϭ⌽Ϫ1͓F Y i͑Y i͔͒,(10) where⌽Ϫ1is the inverse of a standard normal distribution and F is a CDF of a general random variable Y.Equation͑10͒implies that the transformation maintains the CDFs being identical both in the original random space(Y-space͒and the U-space.The MPP is formally deﬁned in the standardized normal space as the minimum distance point on the constraint boundary g(d,X,P)ϭg(d,U X,U P)ϭ0to the origin.The minimum distance ␤is called reliability index.When the First Order ReliabilityMethod͑FORM͓͒28͔is used,the reliability is given by␣ϭProb͕g͑d,X,P͒р0͖ϭ⌽͑␤͒,(11) where⌽is the standard normal distribution function.Finding the MPP and the reliability index is a minimization problem,which usually involves an iterative search process.For details about the MPP based method,refer to͓26͔.In an inverse reliability problem,the required reliability␣is given and the percentile performance corresponding to␣is to be evaluated.Form Eq.͑11͒,the reliability index␤is given by␤ϭ⌽Ϫ1͑␣͒.(12) Note that Eq.͑12͒is applicable for␣у0.5.When␣Ͻ0.5,it becomes␤ϭ⌽Ϫ1͑1Ϫ␣͒.(13) As shown in Fig.4,the MPPIR becomes the common point ͑tangent point͒of a hyper sphere with radius␤͑␤-sphere͒in U-space and the contour of g(U).At this point g(U)reaches its minimum͑or maximum͒.Whether the function is minimum or maximum at the MPPIR depends on to which tail the MPPIR corresponds.When the MPPIR corresponds to the left tail,we have a minimization problem otherwise we have a maximization problem.We will only discuss the maximization,but the same principle can be applied to a minimization problem.Fig.4shows the MPPIR where g(U)is to be maximized.An MPPIR problem is modeled as a maximization problem:ͭmaxmize g͑u͒subject to͑u T u͒1/2ϭ␤(14) Once the MPPIR is identiﬁed,the percentile performance is cal-culated byg␣ϭg͑u M PPIR͒ϭg͑x M PPIR͒,(15) which is the g function evaluated at the MPPIR.Several existing methods can be used to solve Eq.͑14͒,includ-ing optimization techniques͓29͔,the traditional MPP search algo-rithm͓26͔that is based on the concept of the steepest ascent direction͑we will discuss it later in Eq.17͒,the Diagonal Direc-tion Method͓30͔,and the Hybrid Mean Value͑HMV͒Method ͓31͔.Solving Eq.͑14͒using conventional optimization techniques is a generic method but may not be efﬁcient to solve the special type of minimization problem in Eq.͑10͒.Other specialized meth-ods are mostly gradient-based but there is no guarantee of conver-gence.The solution found could be a saddle point or a minimum point instead.Some of the existing MPP or MPPIR search algo-rithms have convergence difﬁculties for non-concave and non-convex problems.It is our goal in this research to develop a new efﬁcient MPPIR search algorithm that can be used for any types of performance functions and is robust in its convergence behavior.4.2The Proposed MPPIR Search Algorithm.In develop-ing an improved MPPIR search algorithm,we aim to improve the performance of the algorithm in two categories:1͒efﬁciency:to ﬁnd the MPPIR with the number of function evaluations as small as possible for any regular͑well-behaved͒functions and2͒ro-bustness:to avoid divergence caused by irregular performance functions.Our proposed algorithm starts from the same vector overlap-ping condition which is used in the traditional search algorithm. Referring to Fig.4for a two dimensional case,the MPPIR is the tangent point of the␤-sphere and the limit state surface in the U-space.At this tangent point,the vector u M PPIR connecting the MPPIR u M PPIR and the origin O should overlap with the gradient ٌg(u M PPIR)of the function g(U).The angle between u k andٌg(u k)is calculated by␥kϭcosϪ1u k•ٌg͑u k͒ʈu kʈ•ʈٌg͑u k͒ʈ(16)At the MPPIR,the angle␥k should be zero.A sufﬁciently small angle␥k is considered as the stopping criterion.To satisfy this condition,the search process starts from the steepest ascent direc-tion and this is what a traditional MPPIR search algorithm does. When this direction leads to a decreasing performance function value due to the irregular function behavior,the second measure—an arc search procedure will be performed.It is called arc search because the search of the MPPIR is along an arc of the ␤-sphere.The arc search can avoid converging to a minimum point or a saddle point.Note that for convenience,the search procedure is illustrated here in a two dimensional space.For higher dimensional problems,a plane,a curve,or a circle dis-cussed for the two dimensional case will be a hyper plane,a hyper surface,or a hyper sphere,respectively.Suppose the current point is u k(k stands for the k th iteration in searching the MPPIR͒.Atﬁrst,the steepest ascent direction ٌg(u k)is used to obtain the new point u kϩ1on the␤-sphere by the following equation͓28͔,u kϩ1ϭ␤ٌg͑u k͒ʈٌg u kʈ(17)Since the feature of the steepest ascent direction ofٌg(u k)is valid locally around u k,there is a need to check the performance function value to see whether there is a progress when using Eq.͑17͒.If g(u kϩ1)Ͼg(u k),there indeed is a progress and the next iteration will follow Eq.͑17͒again.If g(u kϩ1)рg(u k),it indi-cates that u kϩ1is not improved compared with u k.An arc search will then be performed to identify a new u kϩ1that leads to a increasing value of performance.As shown in Fig.5,the arc search is toﬁnd the maximum function value point on the intersection of␤-sphere and the plane determined by the vectors u k andٌg(u k).Apparently,theplane Fig.4Inverse most probable pointJournal of Mechanical Design JULY2004,Vol.126Õ565passes through the origin O and the search path is an arc of the ␤-sphere.Let ␥k denote the angle between the next point u k ϩ1and the current point u k ͑see Fig.5͒.u k ϩ1is expressed by a linearcombination of u k ϩ1and ٌg (u k )as follows.u k ϩ1ϭ␤sin ͑␥k ͒ͩsin ͑␥k Ϫ␦k ͒u k ʈu k ʈϩsin ␦kٌg ͑u k ͒ʈٌg ͑u k ͒ʈͪ(18)The arc search is then formulated as a one-dimensional maxi-mization problem represented byͭFind :The angle ␦k Maximize g ͑u k ϩ1͒ϭgͭ␤sin ͑␥k ͒ͩsin ͑␥k Ϫ␦k ͒u k ʈu k ʈϩsin ␦kٌg ͑u k ͒ʈٌg ͑u k ͒ʈͪͮ(19)Once the optimal angle ␥k is found,the new point u k ϩ1iscalculated by Eq.͑18͒.To further illustrate the procedure of an arc search,the progress of the proposed method is demonstrated in Fig.6for a three-dimensional case.Suppose from the k -th iteration,an arc search is needed and the current point is u k ,the new point u k ϩ1is deter-mined in the plane by the vector u k and vector ٌg (u k ).Geometri-cally,this new point is the tangent point of the projection of ٌg (u k )on ␤-sphere ͑an arc segment ͒and the projection of g (u k )on ␤-sphere ͑contours on ␤-sphere ͒.Obviously,the value of g (U )at u k ϩ1is smaller than the one at u k .Analogously,the same procedure is conducted to ﬁnd points u k ϩ2,u k ϩ3,¯,etc.until the vector u i overlaps with the vector ٌg (u i ).If we let the current point be u k on the ␤-sphere,the search process is summarized as follows:1.Calculate the gradient ٌg (u k )at u k .2.Calculate the angle ␥k between ٌg (u k )and u k using Eq.͑16͒;3.If ␥k р␧,u k is the MPPIR and go to 4,otherwise go to 1.␧is a small angle,for example,0.1°.4.Calculate the percentile performance g (u k )and stop;5.If g (u k )Ͻg (u k Ϫ1),update the pointu k ϩ1ϭ␤ٌg ͑u k ٌ͉͒g ͑u k ͉͒,and k ϭk ϩ1,then go to 1.Otherwise,use Eqs.͑18͒and ͑19͒to perform the arc search to locate the new point u k ϩ1and update k by k ϭk ϩ1.Then,go to 1.To make the search process robust and efﬁcient,the adaptive step size is also employed for the ﬁnite difference derivative evaluations if analytical derivatives are unavailable.The step size in one axis is 1%magnitude of the corresponding component of the current U point.For a ‘‘well-behaved’’performance function,for example,a convex function,the steepest ascent direction method works well and function g increases constantly.In this case the efﬁciency of our method is as good as the traditional method.When function g is convex,or non-concave and non-convex,the arc search in our proposed method guarantees the ascent of the performance func-tion and therefore the convergence.Hence our proposed method is robust to various types of limit-state functions.We will further verify our algorithm through comparative studies.4.3Veriﬁcations of the MPPIR Search Algorithm.Three examples are presented in this paper for the purpose of veriﬁca-tion.In all examples,a ﬁnite difference method is used for deriva-tive evaluations.The efﬁciency and robustness of the proposed algorithm is compared with the traditional MPP search algorithm ͑based on the concept of the steepest ascent direction ͒as well as using the Sequential Quadratic Programming ͑SQP ͒for solving directly Eq.͑16͒.The SQP algorithm is chosen because it is a widely accessible and mature optimization solver.All examples involve functions taken from engineering applications,but the de-tailed background is omitted.Example 1:g ͑X ͒ϭX 2ϩ͑X 1ϩ0.25͒2Ϫ͑X 1ϩ0.25͒3Ϫ͑X 1ϩ0.25͒4Ϫ4(20)where X 1ϳN (0.0,1.0)and X 2ϳN (0.0,1.0);N (␮,␴)stands for a normal distribution with mean ␮and standard deviation ␴.TheFig.5ArcsearchFig.6Arc search procedure566ÕVol.126,JULY 2004Transactions of the ASME。

基于高分辨率航空遥感影像的林分因子智能识别技术研究

第48卷第4期2023年7月㊀林㊀业㊀调㊀查㊀规㊀划Forest Inventory and PlanningVol.48㊀No.4July2023doi:10.3969/j.issn.1671-3168.2023.04.003基于高分辨率航空遥感影像的林分因子智能识别技术研究李琦1,辛亮2,孟陈3(1.上海市林业总站,上海200072;2.上海市测绘院,上海200063;3.景遥信息技术有限公司,上海201109)摘要:森林资源监测的数字化和智能化是未来发展的主要趋势㊂基于高分辨率航空㊁多光谱遥感数据和数字地表模型(DSM)等数据,利用计算机深度学习方法,研究乔木林小班的郁闭度㊁平均树高㊁总株数3项主要林分调查因子的数字化智能提取方法㊂结果表明,郁闭度判读的平均准确率可达到98.6%;平均树高判读的平均准确率可达到90%;株数判读的平均准确率可达到82.36%㊂关键词:智能识别技术;高分辨率航空遥感影像;林分调查因子;自动判读中图分类号:S771.8;TP75;TP18㊀㊀文献标识码:A㊀㊀文章编号:1671-3168(2023)04-0024-04引文格式:李琦,辛亮,孟陈.基于高分辨率航空遥感影像的林分因子智能识别技术研究[J].林业调查规划,2023, 48(4):24-27.doi:10.3969/j.issn.1671-3168.2023.04.003LI Qi,XIN Liang,MENG Chen.Intelligent Recognition Technology of Forest Stand Factors Based on High-resolution Aerial Remote Sensing Images[J].Forest Inventory and Planning,2023,48(4):24-27.doi:10.3969/j.issn.1671-3168.2023.04.003Intelligent Recognition Technology of Forest Stand Factors Based onHigh-resolution Aerial Remote Sensing ImagesLI Qi1,XIN Liang2,MENG Chen3(1.Shanghai Forestry Station,Shanghai200072,China;2.Shanghai Surveying and Mapping Institute,Shanghai200063,China;3.Jingyao(Shanghai)Information Technology Co.,Ltd.,Shanghai201109,China) Abstract:The digitization and intelligence of forest resource monitoring is the main trend in future devel-opment.Based on high-resolution aerial,multispectral remote sensing data,and digital surface model (DSM)data,this paper studied the digital intelligent extraction method for three main forest stand survey factors,namely canopy density,average tree height,and total plant number,in the subcompartment of arboreal forest by using computer deep learning.The results showed that the average accuracy of canopy density interpretation reached98.6%;the average accuracy of average tree height interpretation reached 90%;the average accuracy of plant number interpretation reached82.36%.Key words:intelligent identification technology;high-resolution aerial remote sensing images;forest stand survey factors;automatic interpretation㊀㊀森林是陆地生态系统中的重要主体,为人类提供赖以生存和发展的重要物质基础,对人类生存和经济社会可持续发展起着不可替代的作用[1]㊂上海市是最早进行城市森林规划和建设的城市之收稿日期:2022-03-08;修回日期:2022-06-20.基金项目:上海市绿化和市容管理局科学技术项目(G201209).第一作者:李琦(1969-),男,浙江金华人,高级工程师.主要从事森林资源监测工作.李琦,等:基于高分辨率航空遥感影像的林分因子智能识别技术研究一[2-3]㊂城市森林在固碳[4-5]㊁生物多样性保护[6]和缓解城市热岛[7-8]等方面均具有重要作用㊂森林资源监测是对森林资源的数量㊁质量㊁空间分布㊁利用状况等现状及其动态消长变化进行观测㊁分析和评价的一项林业基础性工作,为实现森林资源科学管理和合理利用提供重要的基础数据保障㊂森林资源监测的主要工作内容之一,即是对森林资源的树种㊁胸径㊁树高㊁郁闭度㊁株数等林分因子状况进行调查㊂传统的人工调查方式存在着诸多不足之处,主要表现在两个方面:(1)大量的监测数据需要调查人员实地获取,工作任务繁重,人力成本高,监测效率较低;(2)监测数据的精度依赖于调查人员的专业素质和工作责任心,主观因素影响较大,监测成果的质量不稳定㊂随着科学技术水平的发展,无人机㊁激光雷达等新装备以及遥感信息自动识别㊁计算机深度学习等新技术已逐步在森林资源监测领域得到重视和广泛研究[9-15]㊂2020年以来,上海市林业总站联合上海市测绘院㊁国家林业和草原局华东调查规划院等多家单位,基于高分辨率航空遥感影像㊁多光谱遥感影像㊁数字地表模型(DSM)等数据,利用计算机深度学习方法,开展了以乔木林小班郁闭度㊁平均树高㊁总株数3项主要林分调查因子为重点研究内容之一的自动估测方法研究,以实现减少外业实地调查工作量,提高监测效率和调查精度,提升上海市森林资源监测智能化水平的目标㊂1研究区域概况上海市地处长江入海口,位于长江三角洲以太湖为中心的碟形洼地东缘,地势低平㊂是一座具有世界影响力的现代化国际大都市㊂据上海市2020年度森林资源监测成果显示,全市森林总面积为117258hm2,森林覆盖率为18.49%㊂其中,乔木林面积为104028hm2,占森林总面积的75.87%㊂上海市森林资源具有典型的城市森林特点,全部为人工林,混交林分占比超过50%,林分组成较为复杂,其功能以生态防护或景观游憩为主,生态公益林面积占比超过80%㊂由于境内水网密布,道路纵横,城镇化水平高等因素影响,森林资源小班的尺度较小,细碎小班多,分布零散㊂全市共有各类森林资源小班约35万个,其中0.5hm2以下小班约28.5万个,占比达81.81%㊂四旁树的分布在全市占有较大比重㊂2研究方法2.1数据源根据上海市森林资源分布特点,本项研究以2020年度森林资源监测成果数据库中乔木林小班为研究对象,基于小班现状分布及其边界范围开展郁闭度㊁平均树高㊁总株数3项主要林分调查因子的自动估测技术研究㊂研究采用的影像数据主要包括上海市2019年11月至2020年2月大飞机DMC III (digital mapping camera)航空遥感影像,分辨率为0.1m,波段为RGBN(红㊁绿㊁蓝㊁近红外)4个波段;以及依据航空摄影测量影像通过空三解算的数字地表模型(DSM)数据(分辨率为0.1m)㊂正射影像和DSM模型经布设于地面的检查点核验,平面㊁高程精度分别为ʃ0.12m和ʃ0.19m㊂辅助数据源为2020年第三季度高景一号卫星影像,分辨率为0.5m㊂2.2技术路线基于现有乔木林小班边界范围,利用本市高精度四波段航空影像提取小班内植被覆盖范围,然后利用DSM数据进行切片处理,剃除低矮植被后提取出乔木分布和冠幅信息,进而获取小班郁闭度㊁平均树高㊁总株数等林分调查因子信息(图1)㊂图1㊀技术路线Fig.1㊀Technical route2.3研究方法2.3.1郁闭度判读郁闭度是指林木冠层的投影面积与该林分林地㊃52㊃第4期林业调查规划总面积的比值㊂郁闭度准确判读的关键在于乔木层林冠范围的准确识别与提取㊂本研究采用归一化植被指数算法进行小班郁闭度的自动判读㊂归一化植被指数是一种常用的植被提取算法,利用归一化植被指数算法可以提取小班内所有的植被信息,然后利用DSM点云数据,通过设置高度阀值剔除草地㊁农用地㊁绿化地表等低矮植被信息,即可获得较为准确的乔木冠层信息㊂由于上海市航空影像的获取时间为冬季,基于光谱信息的冠幅提取方法对落叶树种并不完全适用,判读精度不足㊂因此,本研究中同时利用了植被生长状态较好的夏季卫星影像来进行弥补,由于卫星传感器无法获取立体影像对,基于DSM数据的林冠提取方法并不适用,因此,在已有数据的基础上,通过制作林地样本,利用计算机深度学习技术提取乔木冠幅信息,进而获得了较为真实的小班郁闭度㊂随机选取46个乔木林小班,利用判读数据与人工在航片上区划产生的数据进行精度比对,结果显示,郁闭度判读的平均准确率可达98.6%㊂2.3.2平均树高判读平均树高是反映林分中所有林木高度平均水平的测树因子,是森林资源调查中重要的调查因子之一㊂基于遥感影像开展林分平均树高估测的研究方法很多[13-15],本研究主要利用数字地表模型(DSM)点云技术进行判读㊂DSM是指包含了地表建筑物㊁桥梁和树木等高度的地面高程模型,能够真实地表达地面起伏情况,近年来在城市规划㊁林业等部门已得到较广泛应用㊂由于DSM数据记录了地表不同位置的高程信息,因此对DSM数据进行切片处理可得到不同高度平面上的地物信息㊂本研究通过DSM与数字高程模型(DEM)的差值得到乔木林冠层高度模型(CHM),然后利用判读获取的小班内所有乔木的单木位置,在CHM中自动提取小班内每株乔木的高度数据,最后汇总计算得到小班的平均树高㊂随机选取46个乔木林小班,利用判读数据与小班实地调查数据进行精度比对,结果显示,平均树高判读的平均准确率可达到90%㊂2.3.3株数判读根据遥感影像进行单株定位和识别是本项研究的重点和难点㊂最终研究确定的方法是:将遥感获得的可见光㊁多光谱影像和冠层高度模型(CHM)作为输入源,人工标记常绿㊁落叶和常绿落叶混交3种林分的单株乔木样本,并通过构建全卷积深度学习U-Net神经网络进行株数判读㊂U-net神经网络通过波段组合㊁归一化㊁图像增强㊁交叉熵函数和Adam 优化等算法,能够获得U-net模型最优参数组合,并获得单株乔木分布的概率点阵图,根据概率点阵图可以实现单株乔木识别和位置数据提取,最后根据现状小班边界范围,即可获取小班内乔木总株数㊂随机选取100个乔木林小班,利用判读数据与人工在航片上标记产生的数据进行精度比对,结果显示,株数判读的平均准确率可达到82.36%㊂不同林分类型的株数识别准确率不同,主要表现为常绿林分>混交林分>落叶林分的趋势,其中,常绿林分小班的平均准确率为90.77%,混交林分小班的平均准确率为79.15%,落叶林分小班的平均准确率为69.88%㊂3结㊀论在小班郁闭度㊁平均树高的自动判读研究中,由于研究所使用的航空遥感影像及DSM等数据均具有较高的分辨率,再辅助以夏季卫星影像数据进行判读补充,解决了冬季树木落叶的影响,两项研究内容的自动判读结果均符合小班调查精度要求,基本实现了预期目标,证明了其研究方法的有效性㊂在小班林木总株数的自动判读研究中,由于受航空遥感影像拍摄季节㊁树木落叶等因素影响,以及小班破碎化程度㊁树种组成复杂程度㊁林分郁闭度过高㊁环境影响等因素对判读精度造成的干扰,其判读结果距离小班调查精度要求仍有一定差距,说明其研究方法仍有进一步探索和改进的空间㊂研究结果表明,基于高分辨率航空遥感影像的林分调查因子自动判读技术,能够较为快速㊁客观㊁准确地获取乔木林小班的部分主要林分调查因子,对于提高森林资源监测效率和调查精度具有重要的实践意义㊂由于上海市森林资源存在着林相复杂㊁小班尺度较小且破碎化程度高等特点,目前的试验性研究与未来大范围的实践应用之间仍存在较多不可预见性,因此,本研究工作仍有待进一步深化和拓展㊂本项目的研究成果结合主要树种(组)智能识别㊁主要树种(组)胸径树高关系模型研建等相关研究成果的综合运用,智能识别技术才能在上海市森林资源监管工作中得到广泛应用,提升全市森林资源监测智能化水平,实现森林资源监测技术质的飞越㊂利用本项目判读获取的小班平均树高及其它研㊃62㊃第48卷李琦,等:基于高分辨率航空遥感影像的林分因子智能识别技术研究究获取的胸径树高关系模型,可以换算获得小班平均胸径这项重要林分调查因子,结合判读获取的小班乔木总株数,即可自动估算出小班的林木蓄积量,从而为上海市开展林长制考核以及为实现碳中和碳达峰而开展的相关工作提供快捷而准确的数据支撑㊂随着林长制在上海市的全面推行,加强对森林资源更新变化情况的实时监控必将成为各级政府和林业部门的工作重心之一㊂计算机智能识别技术高效的监测效率及其全面㊁客观㊁准确的判读精度,将在实施森林抚育㊁林相改造㊁林地征占用审批㊁森林资源灾后普查等工作中提供强有力的技术支持㊂参考文献:[1]宋永昌.植被生态学[M].上海:华东师范大学出版社,2001.[2]达良俊,杨同辉,宋永昌.上海城市生态分区与城市森林布局研究[J].林业科学,2004(4):84-88. [3]宋永昌.城市森林研究中的几个问题[J].中国城市林业,2004(1):4-9.[4]高业林.基于3S技术的城市森林碳汇能力研究[D].济南:山东建筑大学,2021.[5]张桂莲.基于遥感估算的上海城市森林碳储量空间分布特征[J].生态环境学报,2021,30(9):1777-1786. [6]宋永昌.城市森林研究中的几个问题[J].中国城市林业,2004(1):4-9.[7]陈朱,陈方敏,朱飞鸽,等.面积与植物群落结构对城市公园气温的影响[J].生态学杂志,2011,30(11):2590-2596.[8]曹璐,胡瀚文,孟宪磊,等.城市地表温度与关键景观要素的关系[J].生态学杂志,2011,30(10):2329-2334.[9]陈芸芝,陈崇成,汪小钦,等.多源数据在森林资源动态变化监测中的应用[J].资源科学,2004(4):146-152.[10]覃先林,李增元,易浩若.高空间分辨率卫星遥感影像树冠信息提取方法研究[J].遥感技术与应用,2005(2):228-232.[11]张巍巍,冯仲科,汪笑安,等.基于TM影像的林木参数提取和树高估测[J].中南林业科技大学学报,2013,33(9):27-31.[12]张志超.多源遥感森林资源二类调查主要林分因子估测关键技术研究及实现[D].西安:西安科技大学,2020.[13]江志向,陈紫璇,练一宁,等.基于航模飞行器摄影数据的森林信息提取方法[J].北京测绘,2017(3):153-157.[14]赵芳.测树因子遥感获取方法研究[D].北京:北京林业大学,2014.[15]韩学锋.基于高分辨率遥感林分调查因子的提取研究[D].福州:福建师范大学,2008.责任编辑:许易琦㊃72㊃第4期。

xx年中级通信工程师考试(互联网技术)上午真题1

xx年中级通信工程师考试（互联网技术）上午真题xx年中级通信工程师考试上午真题卷面总分：分答题时间：120 分钟单项选择题在下列各题的备选项中，请选择1个最符合题意的选项。

1数字通信的即以每秒几百兆比特以上的速度，传输和交换从语音到数据以至图像的各种信息。

A. 大众化 B. 智能化 C. 模拟化 D. 宽带化2通信技术人员行业道德之一是树立服务保障观念，不图名利地位，属于这一内容的有。

A. 质量第一，确保设备完好率 B. 发扬协作精神 C. 不保守技术 D. 树立整体观念3电信职业道德与电信法律 A. 没有区别 B. 有联系有区别 C. 没有联系D. 无联系有区别4电信条例立法的目的之一是为了。

A. 维护电信用户和电信业务经营者的合法利益B. 鼓励竞争C. 实行互联互通D. 提高电信服务质量5电信监管的基本原则是。

A. 普通服务 B. 规范市场秩序 C. 公开、公平、公正D. 保障电信网络和信息安全6电信经营者经营活动的基本原则是。

A.依法经营，保障服务 B.保障服务，遵守商业道德C.公平、公开、公正，接受监督检查D.依法经营，遵守商业道德7经营基础电信业务的条件中，要求经营者必须依法设立的专门从事基础电信业务的公司，在公司的股权结构中，国有股权或股份不得。

A.少于49% B.少于50% C.少于51% D.少于60%8电信服务质量是指。

A. 服务性能质量B. 服务性能质量和网络性能质量C. 业务性能质量和网络性能质量D. 服务质量和业务质量9为了使网络外部性内部化，以增加整体社会福利，解决的办法是实行互联互通和。

A. 增加业务种类 B. 提高市场进入壁垒 C. 降低市场进入壁垒 D. 对用户提供补贴10合同履行的原则有适当履行和原则。

A. 实际履行 B. 全面履行 C. 不适当履行 D. 代偿履行11终端设备是构成通信网必不可少的设备，在下列终端设备中，不是数字终端设备。

A. PSTN电话机 B. ISDN电话机 C. PC机 D. GSM手机12为提高物理线路的使用效率，电信网的传输系统通常都采用多路复用技术，如PSTN网络和GSM网络的局间中继线路传输均采用复用技术。

美国制定毫米级机器人计划应用于重大灾难搜救任务

[8]EL-SEBAIE M G,MELLOR P B.Plastic Instability Conditions when Deep-drawing into a High Pressure Medium [J].International Journal of Mechanical Sciences,1973,15 (6):485-490.[9]YOSSIFON S,TIROSH J,KOCHAVI E.On Suppression of Plastic Buckling in Hydroforming Processes[J].International Journal of Mechanical Sciences,1984,26(6-8):389-402.[10]YANG D Y,NOH T S.An Analysis of Hydroforming of Longitudinally Curved Boxes with Regular Polygonal Cross= section[J].International Journal of Mechanical Sciences, 1990,32(11):877-890.[11]NIKHARE C,WEISS M,HODGSON P D.Buckling in Low Pressure Tube Hydroforming[J].Journal of Manufacturing Processes,2017,28(1):1-10.[12]王仲仁.特种塑性成形[M].北京:机械工业出版社,1995.[13]NIELSEN K B,BRANNBERG N,NILSSON L.Sheet Metal Forming Simulation Using Explicit Finite Element methods[C].EURODYN'93,Trondheim,1993.[14]SALAHSHOOR M,GORJI A,BAKHSHI-JOOYBARI M.The Study of Forming Concave-bottom Cylindrical Parts in Hydroforming Process[J].The International Journal of Advanced Manufacturing Technology,2015,79(5-8):1139-1151.[15]KIM H S,SUMPTION M D,BONG H J,et al.Development of a Multi-scale Simulation Model of Tube Hydroforming for Superconducting RF Cavities[J].Materials Science and Engineering:A,2017,679:104-115.[16]张志超,徐永超,苑世剑.不锈钢外板对5A06铝合金板材液压胀形行为的影响[J].机械工程学报,2016,52(14):40-47.[17]毛献昌,杨连发,黄晖智,等.AZ31B镁合金液压拉深成形件的壁厚分布[J].热加工工艺,2017(5):125-130. [18]包文兵,徐雪峰,戴龙飞,等.等径三通管整体液压成形壁厚分布规律[J].锻压技术,2017,42(4):91-95. [19]陈大勇,徐勇,张士宏,等.圆角结构对新型金属桥塞密封件液压成形性能的影响[J].锻压技术,2017,42(3):123-129.[20]HASHMI M S J.Radial Thickness Distribution around a Hydraulically Bulge Formed Annealed Copper T-joint: Experimental and Theoretical Predictions[C]//DAVIES B J. Proceedings of the Twenty-second International Machine Tool Design and Research Conference,London:Palgrave, 1982:507-516.[21]AHMED M,HASHMI M S J,Estimation of Machine Parameters for Hydraulic Bulge Forming of Tubular Components[J].Journal of Materials Processing Technology, 1997,64(1-3):9-23.[22]LIMB M E,CHAKRABARTY J,GARBER S,et al.The Forming of Axisymmetric and Asymmetric Components from Tube[C]//KOENIGSBERGER F,TOBIAS S A.Proceedings of the Fourteenth International Machine Tool Design and Research Conference,London:Palgrave,1974:799-805. [23]HUTCHINSON M I.BulgeForming of Tubular Components[D].Sheffield:Sheffield City Polytechnic,1988.[24]DAMBORG F F,JENSEN M R.Hydromechanical Deep Drawing[J].Scandinavian Journal of Metallurgy,2000,29 (2):63-70.[25]NAKAMORI T,SHUKUNO K,MANABE K.In-process Controlled Y-shape Tube Hydroforming with High Accurate Built-in Sensors[J].Procedia Engineering,2017,184:43-49.作者简介:顾勇（1983—），男，讲师，博士研究生，主要研究方向为塑性成形工艺；勾治践（1958—），男，教授，主要研究方向为液压成形及模具设计。

实施和优化android上的加密文件系统毕业论文外文翻译

外文原文Implementing and Optimizing an Encryption on AndroidZhaohui Wang, Rahul Murmuria, Angelos StavrouDepartment of Computer ScienceGeorge Mason UniversityFairfax, V A 22030, USA, ,Abstract—The recent surge in popularity of smart handheld devices, including smart-phones and tablets, has given rise to new challenges in protection of Personal Identifiable Information (PII). Indeed, modern mobile devices store PII for applications that span from email to SMS and from social media to location-based services increasing the concerns of the end user’s privacy. Therefore, there is a clear need and expectation for PII data to be protected in the case of loss, theft, or capture of the portable device. In this paper, we present a novel FUSE ( in USErspace) encryption to protect the removable and persistent storage on heterogeneous smart gadget devices running the Android platform. The proposed leverages NIST certified cryptographic algorithms to encrypt the data- at-rest. We present an analysis of the security and performance trade-offs in a wide-range of usage and load scenarios. Using existing known micro benchmarks in devices using encryption without any optimization, we show that encrypted operations can incur negligible overhead for read operations and up to twenty (20) times overhead for write operations for I/Ointensive programs. In addition, we quantified the database transaction performance and we observed a 50% operation time slowdown on average when using encryption. We further explore generic and device specific optimizations and gain 10% to 60% performance for different operations reducing the initial cost of encryption. Finally, we show that our approach is easy to install and configure acrossall Android platforms including mobile phones, tablets, and small notebooks without any user perceivable delay for most of the regular Android applications.Keywords-Smart handheld devices, Full disk encryption, Encrypted , I/O performance.I. BACKGROUND & THREAT MODELA.BackgroundGoogle’s Android is a comprehensive software framework for mobile devices (i.e., smart phones, PDAs), tablet computers and set-top-boxes. The Android operating system includes the system library files, middle-ware, and a set of standard applications for telephony, personal information management, and Internet browsing. The device resources, like the camera, GPS, radio, and Wi-Fi are all controlled through the operating system. Android kernel is based on an enhanced Linux kernel to better address the needs of mobile platforms with improvements on power management, better handling of limited system resources and a special IPC mechanism to isolate the processes. Some of the system libraries included are: a custom C standard library (Bionic), cryptographic (OpenSSL) library, and libraries for media and 2D/3D graphics. The functionality of these libraries are exposed to applications by the Android Application Framework. Many libraries are inherited from open source projects such as WebKit and SQLite. The Android runtime comprises of the Dalvik, a register-based Java virtual machine. Dalvik runs Java code compiled into a dex format, which is optimized for low memory footprint. Everything that runs within the Dalvik environment is considered as an application, which is written in Java. For improved performance, applications can mix native code written in the C language through Java Native Interface (JNI). Both Dalvik and native applications run within the same security environment, contained within the ‘Application Sandbox’. However, native code does not benefit from the Java abstractions (type checking, automated memory management, garbage collection). Table I lists the hardware modules of Nexus S, which is a typical Google branded Android device.Android’s security model differs significantly from the traditional desktopsecurity model [2]. Android applications are treated as mutually distrusting principals; they are isolated from each other and do not have access to each others’ private dat a. Each application runs within their own distinct system identity (Linux user ID and group ID). Therefore, standard Linux kernel facilities for user management is leveraged for enforcing security between applications. Since the Application Sandbox is in the kernel, this security model extends to native code. For applications to use the protected device resources like the GPS, they must request for special permissions for each action in their Manifest file, which is an agreement approved during installation time.Android has adopted SQLite [12] database to store structured data in a private database. SQLite supports standard relational database features and requires only little memory at runtime. SQLite is an Open Source database software library that implements a self-contained, server-less, zeroconfiguration, transactional SQL database engine. Android provides full support for SQLite databases. Any databases you create will be accessible by name to any java class in the application, but not outside the application. The Android SDK includes a sqlite3 database tool that allows you to browse table contents, run SQL commands, and perform other useful functions on SQLite databases. Applications written by 3rd party vendors tend to use these database features extensively in order to store data on internal memory. The databases are stored as single files in the and carry the permissions for only the application that created the be able to access it. Working with databases in Android, however, can be slow due to the necessary I/O.EncFS is a FUSE-based offering encryption on traditional desktop operating systems. FUSE is the supportive library to implement a fully functional in a userspace program [5]. EncFS uses the FUSE library and FUSE kernel module to provide the interface and runs without any special permissions. EncFS runs over an existing base (for example,ext4,yaffs2,vfat) and offers the encrypted . OpenSSL is integrated in EncFS for offering cryptographic primitives. Any data that is written to t he encrypted is encrypted transparently from the user’s perspective and stored onto the base . Reading operations will decrypt the data transparently from thebase and then load it into memory.B.Threat ModelHandheld devices are being manufactured all over the world and millions of devices are being sold every month to the consumer market with increasing expectation for growth and device diversity. The price for each unit ranges from free to eight hundred dollars with or without cellular services. In addition, new smartphone devices are constantly released to the market which results the precipitation of the old models within months of their launch. With the rich set of sensors integrated with these devices, the data collected and generated are extraordin arily sensitive to user’s privacy. Smartphones are therefore data-centric model, where the cheap price of the hardware and the significance of the data stored on the device challenge the traditional security provisions. Due to high churn of new devices it is compelling to create new security solutions that are hardware-agnostic.While the Application Sandbox protects applicationspecific data from other applications on the phone, sensitive data may be leaked accidentally due to improper placement, resale or disposal of the device and its storage media (e.g. removable sdcard). It also can be intentionally exfiltrated by malicious programs via one of the communication channels such as USB, WiFi, Bluetooth, NFC, cellular network etc.Figure 1. Abstraction of Encryption on AndroidFor example, an attacker can compromise a smartphone and gain full control of it by connecting another computing device to it using the USB physical link [33]. Moreover, by simply capturing the smartphones physically, adversaries have access to confidential or even classified data if the owners are the government officials ormilitary personnels. Considering the cheap price of the hardware, the data on the devices are more critical and can cause devastating consequences if not well protected. To protect the secrecy of the data of its entire lifetime, we must have robust techniques to store and delete data while keeping confidentiality.In our threat model, we assume that an adversary is already in control of the device or the bare storage media. The memory-borne attacks and defences are out of the scope of this paper and addressed by related researches in Section II. A robust data encryption infrastructure provided by the operating system can help preserve the confidentiality of all data on the smartphone, given that the adversary cannot obtain the cryptographic key. Furthermore, by destroying the cryptographic key on the smartphone we can make the data practically irrecoverable. Having established a threat model and listed our assumptions, we detail the steps to build encryption on Android in the following sections.V. PERFORMANCEA. ExperimentalSetup For our experiments, we use the Google’s Nexus S smartphone device with Android version 2.3 (codename Gingerbread). The bootloader of the device is unlocked and the device is rooted. The persistent storage on Nexus S smartphones is a 507MB MTD (Memory Technology Device). MTD is neither a block device not a character device, and was designed for flash memory to behave like block devices. In addition to the MTD device, Nexus S has a dedicated MMC (MultiMediaCard, which is also a NAND flash storage technique) device dedicated to system and userdata partition, which is 512MB and 1024MB respectively. Table II provides the MTD device and MMC device partition layout.In order to evaluate this setup for performance, we installed two different types of benchmarking tools. We used the SQLite benchmarking application created by RedLicense Labs - RL Benchmark Sqlite. To better understand finegrained low level operations under different I/O patterns, we use IOzone [7], which is a popular open source micro benchmarking tool. It is to be noted that these tools are both a very good case study for ‘real-use’ as well. RL Benchmark Sqlite behaves as anyapplication that is database-heavy would behave. IOzone uses the direct intensively just like any application would, if it was reading or writing files to the persistant storage. All other applications which run in memory and use the CPU, graphics, GPS or other device drivers are irrelevant for our storage media tests and the presence of encrypted will not affect their performance.IOzone is a benchmark tool [7]. The benchmark generates and measures a variety of and has been widely used in research work for benchmarking various on different platforms. The benchmark tests performance for the generic , such as Read, write, re-read, re-write, read backwards, read strided, fread, fwrite, random read, pread ,mmap, aio read, aio write.IOzone has been ported to many platforms and runs under various operating systems. Here in our paper, we use ARM-Linux version (Android compatible) of latest IOzone available and focus on the encryption overhead. The cache effect is eliminated by cold rebooting the device for each run of IOzone and RL Benchmark Sqlite. The device is fully charged and connected to external USB power while in experiments. We collect the data and plot the average results of the 5 runs in the figures in all the following experiments.A.ThroughputPerformance of EncFS In this section, we present the IOzone performance results for random read and write operations on userdata partition. The benchmark is run for different and for each , with different record lengths. The maximumTable IIISQLITE PERFORMANCE ON GOOGLE NEXUS Sis selected as 4MB due to the observation that 95% of the user data files are smaller than 4MB on a typical Android system.Fig 3 compares the throughput for four typical operations, namely read, random read, write and random write. The IOzone experiments are run on the original ext4 and EncFS with different AES key lengths. Fig 3 shows for read operation, EncFS performs the same with original ext4. However, for random read, write, random write, EncFS only gives 3%, 5%, 4% of the original throughput respectively. Our analysis shows the encryption/decryption contributes the overhead and is the expected trade-off between security and performance. The buffered read in EncFS makes the read operation only incur marginal overhead. However, for random read, the need for the data blocks alignment during decryption results in slower throughput. For different key length, the 256-bits key only incurs additional 10% overhead comparing to 128-bits key for better security. In particular, AES-256 runs 12866KB/s,8915KB/s, 9804KB/s at peak for random read,write and random write respectively while AES-128 runs 14378KB/s, 9808KB/s, 10922KB/s. The performance loss of a longer key length trading better security properties is only marginal to the performance loss of the encryption scheme. Optimizations can compensate such key-length overhead as illustrated in Section V-D. Based on this observation, AES-256 is recommended and used as default in the following subsection unless otherwise mentioned explicitly.Similarly, sdcard partition gives the identical pattern with slightly different value. Due to the fact that the sdcard partition shares the same underlying physical MMC device with userdata partition as listed in Table II, our experiment results demonstrates the original vfat performs 16% faster than ext4 for read and random read operation while ext4 outperforms vfat 80% and 5% for write and random write operations respectively. However, comparing different is out of our focus in this paper. We observed different throughput values and overhead patterns on other devices such as Nexus One, HTC Desire and Dell Streak which use a removable sdcard as separate physical medium to internal NAND device. Both AES-128 and AES-256 throughput on sdcard are statistically identical to the ones on userdata partition given a 95% confidence interval. Such results show that the scheme of encryption in EncFS(e.g. internal data block size, key length) and its FUSE IO primitives are the bottleneck of the performance regardless of the underlying . We suggest corresponding optimizations in Section V-D.In addition to the basic I/O operations, we look at the read operation in detail under different record size before and after encryption. In particular, we plot the 3D surface view and contour view. In the 3D surface graph, the x-axis is the record size, the y-axis is the throughput in Kilobytes per second, and the z-axis is the . The contour view presents the distribution of the throughput across different record sizes and . In a sense, this is a top-view of the 3D surface graph. Figure 4 and 5 show the throughput when IOzone read partial of the the beginning. Figure 4 shows the default ext4 in Android 2.3 favors bigger record size and for better throughput. The performance peak centers in the top-right corner in the contour view of the 3-D graph. However, after placing EncFS, the performance spike shifts to the diagonal where the record size equals to . This is an interesting yet expected result because of the internal alignment of the in decryption.To better understand the performance of our encryption under Android’s SQLite IO access pattern, we present the database transactions benchmark in the next subsection, which is more related to the users’ experiences.C.SQLite Performance BenchmarkingIn addition to the IOzone micro benchmark results in last subsection, we measure the time for various typical database transactions using the RL Benchmark SQLite Performance Application in the Android market [11]. Table III groups the read and write operations and lists the results in detail.We consider that random read and write is a fair representation of database I/O operations in our scenario. This is due to the fact that for SQLite, the database of one or more pages. All reads from and writes to the database at a page boundary and all reads/writes are an integer number of pages in size. Since the exact page is managed by the database engine, only observe random I/O operations.After incorporating the encryption , the database-transactions-intensive apps slows down from 81.68 seconds to 128.66 seconds for the list of operations as described in the Table III. The read operations reflected by select database transactions shows the consistent results with IOzone result: the EncFS buffers help the performance. However, any write operations resulting from insert, update, or drop database transactions will incur 3% to 401% overhead. The overall overhead is 58%. This is the trade-off between security and performance.中文翻译实施和优化android上的加密文件系统王朝晖，拉胡尔Murmuria，安吉罗斯Stavrou计算机科学系乔治·梅森大学费尔法克斯，VA 22030，USA，，摘要：比来激增的智能手持设备，包孕智能手机和平板电脑的普及，已经引起了在庇护个人身份信息（PII），新的挑战。

基于深度信念网络的苹果霉心病病害程度无损检测

基于深度信念网络的苹果霉心病病害程度无损检测周兆永1,2，何东健1,*，张海辉1，雷雨1，苏东1，陈克涛1（1.西北农林科技大学机械与电子工程学院，陕西杨凌 712100；2.西北农林科技大学网络与教育技术中心，陕西杨凌 712100）摘要：针对现有霉心病无损检测只能检测出有无病害，无法对病害程度进行判断的问题，研究并提出一种基于深度信念网络（deep belief net，DBN）的无监督检测模型。

该模型由多层限制玻尔兹曼机（restricted Boltzmann machine，RBM）网络和1层反向传播（back propagation，BP）神经网络组成，RBM网络实现最优特征向量映射，输出的特征向量由BP神经网络对霉心病病害程度分类。

对225 个苹果样本在波长200～1 025 nm获取其透射光谱后，根据腐烂面积占横截面比例将霉心病害程度分为健康、轻度、中度和重度4 种，分别用150 个和75 个样本作为训练集和测试集，以全光谱数据和基于连续投影算法提取的特征波长数据为输入构建病害程度判别模型，并比较DBN模型与偏最小二乘判别分析、BP神经网络和支持向量机模型的识别效果，实验结果表明，DBN模型病害判别准确率达到88.00%，具有较好的识别效果。

关键词：苹果霉心病；病害程度；透射光谱；深度信念网络（DBN）；限制玻尔兹曼机（RBM）Non-Destructive Detection of Moldy Core in Apple Fruit Based on Deep Belief NetworkZHOU Zhaoyong1,2, HE Dongjian1,*, ZHANG Haihui1, LEI Yu1, SU Dong1, CHEN Ketao1(1. College of Mechanical and Electronic Engineering, Northwest A&F University, Yangling 712100, China;2. Network and Education Technology Center, Northwest A&F University, Yangling 712100, China)Abstract: Apple moldy core is a major disease affecting the internal quality of apple fruit. However, due to the lack of effective means to accurately detect moldy core in apple fruits, detection of moldy core in apples has become a major problem to be solved in the apple industry. To date, there have been no reports on the use of spectroscopy for distinguishing various degrees of moldy core decay in apple fruits. The objective of this study was to develop a non-destructive method for the detection of various degrees of moldy core decay in apple fruits using near infrared transmittance spectroscopy, successive projections algorithm (SPA), and multi-class classiﬁcation algorithms partial least square-discriminant analysis (PLS-DA), back propagation artiﬁcial neural network (BP-ANN), support vector machine (SVM) and deep belief network (DBN). For developing a model to determine the degree of moldy core in apples, 225 samples were selected including a training set of 150 samples and a test set of 75 samples. The model consisted of several layers of restricted Boltzmann machine (RBM) network, which achieved eigenvector projection, and one layer of BP network, which allowed the classiﬁcation of various degrees of moldy core based on the output eigenvector. The S d value was calculated by dividing the lesion area by the total cross-sectional area. It was proposed that S d = 0, 0 < S d ≤ 10%, 10% < S d < 30% and S d ≥ 30 indicated health, mild, moderate and severe degrees, respectively. The classiﬁcation accuracy of the DBN model was 88.00%, suggesting good performance of the model, and it was compared with those of the BP-ANN and SVM models.Key words: moldy core in apples; degree of disease; transmittance spectroscopy; deep belief network (DBN); restricted Boltzmann machine (RBM)DOI:10.7506/spkx1002-6630-201714046中图分类号：S24 文献标志码：A 文章编号：1002-6630（2017）14-0297-07收稿日期：2016-09-12基金项目：国家高技术研究发展计划（863计划）项目（2013AA10230402）；国家自然科学基金面上项目（61473235）；陕西省重大农技推广服务试点项目（2016XXPT-05）作者简介：周兆永（1980—），男，工程师，博士研究生，主要从事智能化检测、模式识别研究。

遥感类英文期刊、与GIS相关的SCI(EI)期刊

遥感类英文期刊、与GIS相关的SCI（EI）期刊遥感类英文期刊[1]. REMOTE SENSING OF ENVIRONMENTISSN： 0034-4257版本： SCI-CDE出版频率： Monthly出版社： ELSEVIER SCIENCE INC, 360 PARK AVE SOUTH, NEW YORK, NY, 10010-1710出版社网址：http://www.elsevier.nl/期刊网址：http://www.elsevier.nl/inca/publications/store/5/0/5/7/3/3/index.htt影响因子： 1.697(2001),1.992(2002)主题范畴： REMOTE SENSING; ENVIRONMENTAL SCIENCES; IMAGING SCIENCE & PHOTOGRAPHIC TECHNOLOGY[2]. International Journal of Remote Sensing国际遥感杂志,这是英国出版的一本专业遥感杂志，由Taylor & Francis出版社出版，我看得不多，不好多说，大家可以到/上查该杂志目录和摘要，全文可以到系资料室复印。

Taylor&Francis注册邮箱为cumtlp#.[3]. PHOTOGRAMMETRIC ENGINEERING AND REMOTE SENSINGISSN： 0099-1112版本： SCI-CDE出版频率： Monthly出版社： AMER SOC PHOTOGRAMMETRY, 5410 GROSVENOR LANE, SUITE 210, BETHESDA, MD, 20814-2160出版社网址：/期刊网址：/publications.html影响因子： 0.841(2001);1.176(2002)主题范畴： GEOGRAPHY, PHYSICAL; GEOSCIENCES, MULTIDISCIPLINARY; REMOTE SENSING; PATHOLOGY美国摄影测量与遥感协会的会刊，摄影测量工程与遥感(Photogrametry Engineering and remote sensing)杂志,网址，网上可看论文摘要，全文可去系图书室复印，该杂志南大很全，过去50年的都有,同时你可以发现测绘学报、遥感学报、武测学报甚至国外的文章中参考文献许多来自该刊，在2000年SCI地学类杂志排名中该刊名列第三，相当地不容易，南大校友美国 berkeley加州大学副教授宫鹏就曾三次获得该刊的年度最佳论文奖，由此奠定其在国际遥感学界的地位。

非均质秸秆纤维复合材料保险杠蒙皮刚度分析

第14卷第3期精密成形工程刘军舰1，胡豪胜2a,2b，周磊1，李伟1,2a（1. 上汽通用五菱汽车股份有限公司，广西柳州 545007；2. 武汉理工大学 a. 现代汽车零部件技术湖北省重点实验室；b. 汽车零部件技术湖北省协同创新中心，武汉 430070）摘要：目的研究非均质秸秆纤维复合材料保险杠蒙皮的刚度性能。

方法采用试验与模拟分析的方法，通过共混挤出与化学发泡注塑工艺制备微发泡秸秆纤维/聚丙烯（SF/PP）复合材料试样，通过试验测试非均质结构试样的力学性能与微观结构，通过有限元分析手段建立非均质微发泡秸秆纤维/PP复合材料结构分析模型，并分析非均质材料保险杠蒙皮的刚度性能。

结果微发泡秸秆纤维/聚丙烯（SF/PP）复合材料的微观结构有明显的“三明治”结构特点，秸秆纤维主要分布在外皮层，泡孔主要分布在芯层。

将非均质秸秆纤维复合材料保险杠蒙皮近似为3层复合板结构，建模的刚度分析结果与试验测试相差约6%。

结论非均质秸秆纤维复合材料汽车注塑件可近似为3层复合板结构进行数值分析，简化了分析过程，研究结果可用于指导产品性能评估，提高产品开发效率。

关键词：微发泡；植物纤维复合材料；非均质；力学性能DOI：10.3969/j.issn.1674-6457.2022.03.014中图分类号：U465.4 文献标识码：A 文章编号：1674-6457(2022)03-0107-09. All Rights Reserved.Stiffness of Heterogeneous Bumper Fascia Made by Straw Fiber CompositesLIU Jun-jian1, HU Hao-sheng2a,2b, ZHOU Lei1, LI Wei1,2a(1. SAIC GM Wuling Automobile Co., Ltd., Liuzhou 545007, China; 2. a. Hubei Key Laboratory of Advanced Technology forAutomotive Components; b. Hubei Collaborative Innovation Center for Automotive ComponentsTechnology, Wuhan University of Technology, Wuhan 430070, China)ABSTRACT: The work aims to research the stiffness of bumper fascia made of micro foamed straw fiber/polypropylene (SF/PP)composites. Micro foamed SF/PP composites were prepared by blending extrusion and chemical foaming injection processes.The mechanical properties and microstructure of heterogeneous samples were tested by experiment. A structure analysis modelof heterogeneous micro foamed straw fiber/PP composite was established through finite element analysis. The stiffness per-formance of the composite bumper fascia was analyzed. The results showed that the microstructure of micro foamed SF/PPcomposites had obvious "sandwich" structure characteristics. Straw fibers were mainly distributed in the outer skin layer andbubbles were mainly distributed in the core layer. The difference between the analysis results and the experimental test wasabout 6%. The automobile injection parts of heterogeneous straw fiber composite can be approximated to a three-layer compos-ite plate structure for numerical analysis, which simplifies the analysis process. The research results can be used to guide productperformance evaluation and improve product development efficiency.KEY WORDS: micro foamed; plant fiber composite; heterogeneous; mechanical property收稿日期：2021-09-02基金项目：国家自然科学基金（51605356）；中央高校基本科研业务费专项资金（WUT 2019Ⅲ116CG）作者简介：刘军舰（1982—），男，硕士，高级工程师，主要研究方向为汽车零部件先进制造。

1、下载文档前请自行甄别文档内容的完整性，平台不提供额外的编辑、内容补充、找答案等附加服务。
2、"仅部分预览"的文档,不可在线预览部分如存在完整性等问题,可反馈申请退款(可完整预览的文档不适用该条件!)。
3、如文档侵犯您的权益，请联系客服反馈,我们会尽快为您处理(人工客服工作时间：9:00-18:30)。

The Design of SMS Based HeterogeneousMobile BotnetGuining Geng, Guoai Xu, Miao Zhang and Yanhui GuoInformation Security Center,Beijing University of Posts and Telecommunications, Beijing, Chinagengguining@, {xga, zhangmiao, yhguo }@Guang YangChina Information Technology Security Evaluation Center, Beijing, Chinasunwina@Wei CuiInformation Center of Ministry of Science and Technology of the People's Republic of Chinacuiw@Abstract—Botnets have become one of the most serious security threats to the traditional Internet world. Although the mobile botnets have not yet caused major outbreaks worldwide in cellular network, but most of the traditional botnet experience can be transferred to mobile botnet on mobile devices, so mobile botnet may evolve faster since techniques are already explored. From the theoretical work of some researchers and the reports of security companies, we can see that the mobile botnet attacks and trends are quite real. In this paper, we proposed a SMS based heterogeneous mobile botnet, and shown how SMS based C&C channel structure can be exploited by mobile botnets. At last, we give the analysis of connectivity, security and robustness evaluation of our model.Index Terms—heterogeneous, mobile botnet, SMS, C&C, robustnessI.I NTRODUCTIONTraditional botnets have become one of the most serious security threats to the Internet. The word “bot” means that those victims controlled by attacker, and it derives from the word “robot”. A bot master can control a large scale of bots at different locations to initiate attack, and due to the complexity of the internet, it can be hardly trace back, and lots of researchers have done remarkable studies about the traditional botnet, such as botnet threats[1], botnet model[2], control strategies[3] and botnet detection[4].Compared with the evolution of traditional botnet on the Internet world, mobile botnet are 7-8 years[5] behind. Although the mobile botnets have not yet caused major outbreaks worldwide in cellular network, but most of the traditional botnet experience can be transferred to mobile botnet. Norman[5] predicts that malware on mobile device(smart phones) will evolve faster since techniques are already explored.The early malware can only perform one or two tasks, like Cabir. Cabir was detected in 2004, it is the earliest mobile botnet we ever known.But in 2009, the situation has changed dramatically. Mobile bots can connect back to a malicious bot server and transfer valuable information of the infected device, like SymbOS.Exy.C[6], Ikee.B[7,8]. According to the report of Symantec[6] on 13 July 2009, SymbOS.Exy.C may the ﬁrst bot on Symbian OS. It is a worm similar to other worms that made for Symbian OS, but the difference is that the bot node tries to contact a malicious bot server and transfer valuable information, such as the phone types, International Mobile Equipment Identity (IMEI) and International Mobile Subscriber Identity (IMSI)[9]. Symantec mentions that this may be the first true botnet occurred on a cellular mobile device.BBOS_ZITMO.B[10] was detected in 2011. It receives commands via SMS, steal users information by forwarding SMS messages to a set/predefined admin phone number, monitors incoming calls and SMS. The most important is that it also has a stealth mechanism that prevents being seen as an installed app.Figure1 illustrates the simplified typical infection cycle of an SMS based mobile botnet. There are mainly 4 steps in the botnet operation [11]. Vulnerable smart phones could be ﬁrst infected by bots in the original botnet. The existed techniques (smart phone OS vulnerabilities, Trojan horses, worms, etc) were used during the infection period. After the infection, the infected node connects to the bot server to join in the origin botner. All the bot servers are organized as the C&C network and controlled by bot master. Bot master use the C&C network to issue commands, and control the whole botnet. According to our analysis, the bots in the same tier or sub network could launch the corresponding attack after receiving the command. Authorization is achieved via a channel password.Figure 1 Typical infection cycle Consider the potential threats of the mobile botnet to the cellular network. Researchers have done some remarkable research about the attacks on cellular network, such as DoS attack [12], C&C channel study[13], paging channel overloads [14]. In this paper, we proposed the SMS based heterogeneous mobile botnet. The C&C channel is SMS based heterogeneous multi-tree structured network. That is, all C&C commands are transfered via SMS messages since SMS is available to almost mobile phones. We use heterogeneous multi-tree topology to enhance the scalability and robustness of the botnet. We also made all the bot lists and some of the important commands encrypted. Our research has the following main contributions: • We proposed an improved SMS basedheterogeneous mobile botnet. The heterogeneousbot nodes and networks structure have made themobile botnet more efficient to communicate andmore secure.• Contrast with the traditional botnet of Internetworld, we give the definition of mobile botnet andillustrate the characteristics of the mobile botnet.• We introduced node degree threshold k 〈〉 andmobile botnet height H to improve the security ofthe C&C channel. The improved C&C channelraises the bar for the countermeasure of mobilebotnet community.• Through the design of the eviction and replacement mechanism of failed or recovered botserver node, the robustness of C&C channel wasenhanced.The remainder of the paper is organized as follows.Section II shows the characteristics of cellular bonnets.Section III illustrates the mobile botnet attacks. SectionIV proposed our SMS based heterogeneous cellularmobile botnet. Section V discussed our model from theaspect of connectivity, security and robustness. The paperconcludes and future work with Section VI.II. THE CHARACTERISTICS OF M OBILE BOTNETSIn this paper we define the mobile botnet as follows:Definition 1 mobile botnet: Consists of a network with compromised smart phones, controlled by attacker (“bot master”) thorough a command and control(“C&C”) network for malicious purposes.The mobile botnets are different from the traditional botnets of Internet world. Since the mobile botnets are mostly communicated between mobile devices (smart phones), it has the following characteristics.• Limited by the power resource. The mobile devices such as smart phones are different from PC, its run time is limited due to the use of battery. • The communication costs problems. The communications of mobile botnets will consume the limited power resource, network traffic and especially the phone charge. The communications of the mobile botnets will lead to the cost of theowner, and a significant rise of the phone charge will result in the investigation of the cause and thus may lead to the exposure of the culluar bot. The SMS based mobile botnets, depending on the type of mobile phone contract, SMS messages can be completely free or charging very few money. • The connectivity changes constantly, even unstable. The connectivity may be both affected by physical environment or personal factors. It can be affected by the networks around the mobile phone owner, the action of the mobile phone owner, such as the user is in the tunnel or turn off the mobile phone during the bed time. As shown in Table 1, the connectivity of bots can be changed constantly during the daily life.Table 1 THE C ONNECTIVITY AND T IME I NTERVAL Connectivity Time Intervals WiFi Morning (at home) GSM/EDGE/3G Day time (at work/school)WiFi Day time (at Starbucks) No signal Day time (at wild ) WiFi Evening (at home) Turn off the mobile phone Night (bed time) • Lack of IP address. The lack of IP address maycause the problem of indirect connect. Due to thelack of IP address, most mobile phones are usingNAT gateway and thus the devices are not directlyreachable, so the traditional P2P based C&Cnetwork may not suit for mobile botnet. • The diversity of operating system of smart phone.The design of mobile botnet has to consider thediversity of the OS platform of smart phone.• Mobile botnets are hard to be detected. Evidencesindicate that, mobile botnets are becoming moreand more sophisticated[10]. The valuablemessages can be forwarded by SMS messages to apredefined server device, and the SMS messagesare deleted immediately after they were forwardedby mobile botnets, so it is hard to be detected. It also has a stealth mechanism that prevents beingdetected as an installed app..III. MOBILE BOTNET ATTACKA. Information LeakageOne of the main targets of the mobile botnet is to retrieve sensitive information from the victims. The mobile bot can quickly scanning the host node for signiﬁcant corporate or ﬁnancial information, such as usernames and passwords, address list and text messages.B. DoS AttackBecause most of the functionality of cellular network rely on the availability and proper functioning of HLRs(Home Location Register), so the DoS attack could block the legitimated users of a local cellular network from sending or receiving text messages and calls[9,12,15,17].In the practical circumstances, a bot master of a mobile botnet could control the compromised mobile phones to overwhelm a specific HLR with a large volume of traffic. Through the DoS attack, it will affect all the legitimated users who rely on the same HLR, their requests will be dropped.By overloading (red arrow) the HLR of a local central component of the cellular network, the network will unable to server legitimated traffic (blue arrow) of very large geographic regions. Figure 2 shows how a cellularFigure 2 A cellular network DoS attackC. Charge lossThere exist some service which the smart phones can give money to charity organizations[17]. If the mart phone called or sent a text message to the specific service number, then the subscriber will pays a preset amount of money. The bot master can also creates its own service number and programs all the bots to call or sent a text message to the specific service number. Of course, the price should be low, so the subscribers would not notice and be suspicious about the extra charges.IV.SMS BASED HETEROGENEOUS MOBILE BOTNETDESIGNAs we known, the three main components of botnet are vectors, C&C channel and network topology. The vector solved the problem of how to spread bot code, and can be used in propagation period; C&C channel is used to issue command and control messages; the network topology of botnet is to manage the botnet and enhance the security and stealth of botnets.Considering the problems that have been encountered during the design period of traditional botnets and mobile botnets, we think the bot master should consider the following challenges:1. How to transmit command and control messages and control traffic flow.2. How to prevent the proposed mobile botnet from being detected by defenders.3. How to monitor the status of bot nodes.4. How to enhance the robustness of mobile botnet even some bot nodes have been removed by the defenders. In this section, we will propose the SMS based heterogeneous mobile botnet, and briefly overview our mobile botnet and all type of nodes and then focus on the propogation, network topology, command and control mechanism, node replacement mechanism of our mobile botnet. The designed mobile botnet model is shown inFigure 3 The simplified SMS-based heterogeneous mobile botnet A. Node TypesThe infected node can be classified by the difference of performance, battery resource, connectivity and networks around it, so they can be divided into several types of bot nodes. As is illustrated in Figure 3, our mobile botnet is made up of bot master, collection node, bot servers, region bot servers and a certain amount of bot nodes. The topology of our mobile botnet was designed multi-tree structured from top to down, but the topology of the same tier nodes such as bot server tier was designed P2P structured. So, the heterogeneous exist both in bot nodes types and network topology.Bot master node: Bot master controls all the nodes of the mobile botnet. The bot master has direct contact with bot servers. It has the bot list of all the botnet, and it stores all the node’s information, like user name, phone number, international mobile subscriber identity (IMSI), International Mobile Equipment Identity(IMEI). Unless an emergency situation, bot master will not communicate with bot node directly.Collection node:It receives the valuable information from all the nodes of the botnet. The stored information can be fetched by bot master. Before it receives the information, the node that sends information must be authenticated.Bot server node:Bot Server node is one of the key nodes of our model. It has two main functions, search and forward, it has 'k direct links with region bot servers, and each of them controls a sub network. It has the bot list that includes all the nodes it communicated with, and the bot list is encrypted.Region bot server node: Region bot server node is another key node of our model. It is both boy server and bot node. It receives commands from bot server and forwards to the bots of its sub network. It executes the commands and transfer valuable information to the collection node. Also its bot list was encrypted.Bot node: Bot node is the leaf node of ourmobile botnet. It receives commands from region bot server, and executes the commands.The comparison of communication and bot list between the five types of nodes that we have stated is shown in table2.Table 2THE COMMUNICATION OF MODEL NODESNode type Bot master Collection node Bot server Region bot server BotCommu. Node list Commu. Node list Commu.Node list Commu.Node list Commu.NodelistBot master yes yes yes yes yes Collection node yes yes yes yes yes Bot server yes yes yes yes yes yes Region bot serveryes yes yes yes yes yes Bots yes yes yes yes yes yes yes B. PropogationWe divide the propagation phase into three steps. In the first step, the bot master exploit the operatingsystem and configuration vulnerabilities to compromisethe mobile devices, install the bot software and collecting valuable information. After the bot software was installed, we assume the bot node has a stealth mechanism that prevents being seen as an installed app, and the messages that used to collect valuable information were deleted.In the second step, the bot master choose bot server nodes from the compromised nodes as the first tier nodesaccording to the following conditions, the energyresource, the region where it belongs to and theconnectivity. The rest of the nodes that are not chosen as the bot servers, assigned to the bot servers according the regions, as the second tier nodes.The third step, the nodes of the existed botnets continued infecting mobile devices under the control ofthe bot master. According the expansion of the mobilebotnet, the number of tier will grow.C. Network TopologyDefinition 2 Botnet degree: The degree of our mobile botnet is K . It is the maximum value of ,i k i N 〈〉∈. Definition 3 Botnet height: The height of our mobile botnet is H . It is the number of tiers of the model.Definition 4 A tomic network: The atomic network iscomposed of region bot node and 'k bot nodes, it is the smallest bot network that was control by region bot server at the lowest tier.Definition 5 'k : we define 'k is the number of lower tier nodes that communicate with higher tier node n i .In general, our proposed mobile botnet is made up of 1 bot master, 'k bot servers, 'k region bot servers, and'k sub networks.The maximum degree of bot server n i is '2k k 〈〉=, in Figure 3, 'k =3, k 〈〉=6. For mobile botnets with largescale bot nodes, it makes sense to limit the degree ofbotnet K under a threshold. So, unless the bot nodes arehigh capacity server, bot masters should keepk 〈〉small[14]. As we said, the topology of our model is heterogeneous multi-tree structured. From top to down, the model is a multi-tree structure network. The bot server tier and region bot server tiers are the P2P structured networks. The degree of the model is K , theheight of the model is H . The degree of node n i is i k 〈〉, K is the maximum number of the degree of all the nodes, 12max(,,...,)N K k k k =〈〉〈〉〈〉. Our botnet(see Figure3) can be divided intoH -1 tiers and 'k subnets. The lowest tier network contains 'k ('k =3) subnet, and eachsubnet is composed of 'k P2P structured atomic networks, as is shown in Figure 4.The topology between bot server nodes is P2P structured network (tier 1in Figure 3), the distances between them are all 1 hop. Also the topology between region bot server nodes and bot nodes are P2P structurednetwork. This will enhance the communication efficiency between the mentioned bot nodes,.D. Command and Control NetworkIn our mobile botnet model, we use SMS as the C&C channel. Most of the traditional botnets are using centralized IRC, HTTP protocol and P2P based C&C channel, and all of these channels are IP-based C&C delivery. Unlike the PC world, even with the edge or 3G networks, the smart mobile devices (smart phones) still can not establish stable IP based connections with each other. Given this limitation, in our work we use SMS as our un-centralized and non-ip-based C&C channel.The advantages of SMS based C&C channel can be listed as follows.First, it is popular. Most of mobile phone subscribers are sending and receiving text messages to each other. Second, it is easy to use. Malicious content can be hidden in message, the command and control messages can be disguised as spam-looking messages.Third, it has high connectivity. Even a phone has signal problem or power off, the SMS messages that send to it will eventually arrive to it, because the SMS messages can be stored in a service center and delivered once the signal becomes available or turned back on[19]. Our mobile botnet can be considered as “PUSH” based botnet. The “PUSH” based botnet deliver commands form bot server to bot nodes whenever the bot master wants, and we call this process “PUSH”. On the contrary, the traditional botnet [20] and [21] are the “PULL” based botnets, the bot nodes of them periodically communicate with bot server to get commands. But the “PULL” based botnets have some drawbacks, such as generate additional information, have unexpected latency in command delivery. G. Gu et al [18] has designed an effective detection schemes for it. All of the above statements made the “PULL” based botnets unattractive in practice. The “PUSH” based delivery does not generate additional information, and have little latency in command delivery, in this paper, we design our model as “PUSH” based botnet.As is illustrated in Figure 3, The basic design idea for our SMS-based mobile botnet is to use a heterogeneous multi-tree network as the Command and Control(C&C) networks, the network is like a commands channel, all the nodes of network(black nodes) are served as bot server. The C&C network of our proposed model is composed of bot master, bot server and region bot server.Like Kademlia[22], the absence of authentication mechanism means that anyone can insert values to the commands, the defender may use this launch index poisoning attacks and the C&C may be disrupted. So before communication the nodes should be authenticated, and the important command should be encrypted. For the efficiency of C&C, only critical commands issued by the bot master that order bots to execute malicious tasks such as “transmit valuable information” are encrypted, while commands for P2P communication purposes such as“search node” are only disguised without encryption[19].Figure 5 The C&C networkAccording to the bot list of it, bot server node performs two main function, search and forward. Some of the important commands are encrypted, and the encrypted commands are pushed from top tier nodes to lower tier nodes, from the bot master to bot server node, to region bot server node and bot node respectively. All of the bot lists are encrypted, once the key node was captured, the defender will not get the other nodes that communicate with it. And, also it cannot trace the bot master.E . Node repalcement mechanismAt some special circumstances, bot node type may need to transfered from each other. Like one of the bot server node is failed or recovered.The bot server nodes and region bot server nodes are the key nodes of our model, so they must be replaced by other nodes if they are out of function, such as failed or recovered. As is shown in Figure 6, at some circumstances, bot node type may need to transfer from each other. Since bot master has the bot list of the failed key node, by sending a message that containing the failed node’s bot list and communication keys to the new key node and that will make it reconnected to the bots.Figure 6 The transfer between three type of nodes1) Node EvictionAssumption1: we assume that the failed or recovered nodes can be detected by bot master.Assumption2: we assume that all bot node have a self destroy mechanism to delete the bot list that stored in it.Table 3T HE LIST OF NOTATIONSNotation Description K BB Keys between bot server and bot master K BS Keys between bot serversK BRKeys between bot server and region botserverFirst, the bot master detect the failed or recovered node n i , then according to phone number of n i that bot master stored in it to refresh the according keys, such as K BB , K BS , K BR. Second, the bot master sends '1k − messages that contain the new keys K BS to all the other bot servers,and send 'k messages that contain K BR to the region bot servers, which n i communicate with. Third, the phone number of n i was deleted from the bot list of bot master, bot server and region bot server. After the eviction period, the failed or recovered node n i can not communicate with the bot nodes that we have mentioned, and also the messages that send to them werediscarded automatically. The eviction procedure isillustrated in Figure 7.2) Node ReplacementAt the replacement period, first, bot master choose a relatively powerful node from low tier bot nodes or region bot server as the new bot server according to standers that we have stated in propagation section. Second, the bot master send 1 message that contains the new keys K BB , K BS , K BR and bot list to the new assigned bot server. Third, the new bot serner need to send '1k − messages to establish communication with the rest botservers and 'kmessages to establish communication with the region bot servers that were communicated with the captured bot server, so the total number of messages during the replacement period is 2'1k −. The replaceprocedure is illustrated in Figure8.Figure 8 The replacement of bot server node V . EVALUATION In this section, we would like to analyze our proposed mobile botnet from the aspect of connectivity, security, and robustness. A. ConnectivityIn this section, we will discuss the connectivity of our heterogeneous botnet with different scale of node, and compare the connectivity with the P2P structured botnet. We will look at the 200-node botnet first and then the 1000-node and 2000-node botnet. In the 200-node botnet,when 'k =4 and 'k =5, all the nodes in the botnet can be reached in 4 hops, and when 'k =3, all the nodes in the botnet can be reached in 5 hops, as is shown in Figure 8. In the 1000-node botnet, as is illustrated in Figure 9, all the nodes can be reached in 5 hops when 'k =4 and 'k =5, and when 'k =3, all the nodes can be reached in 6 hops. Figure 10 shows the 2000-node botnet, all the nodes will be reached in 5, 6, 7 hops respectively.From Figure 9, 10, 11, we can see that the larger of 'k , the less hops mobile botnet needs to reache all the bot node of it, when the number of hops larger than 3, the connectivity will rise dramatically.Figure 9 The relationship between connectivity and k’, when N=200Figure 10 The relationship between connectivity and k’, when N=1000Figure 11 The relationship between connectivity and k’, when N=2000We compare our proposed botnet with the P2P structured botnet. In the P2P structured botnet, theaverage path length between any two nodes can be 1 hop(a)The shortest path length (b)The longest path lengthFigure 12 the path length of P2P structured BotnetAs is illustrated in Figure 11(a), in the P2P structured botnet, the commands from Bot 0 can be directly forward to Bot 1, Bot 2, Bot 3, Bot 4, Bot 5. So the mean path length is11112_11111N N ii i P Pstructured lPL N N −−=====−−∑∑(1)In Figure 8(b), the commands from Bot 0 can finally forward to Bot 5, through Bot 1, Bot 2, Bot 3 and Bot 4. So the mean path length is11112_2112N N ii i P P structured l iNPL N N −−=====−−∑∑ (2)According to Formula (1) and (2), the mean path length of the P2P structured botnet is212224P Pstructured NN Mean ++==(3)So the mean connection hops between two nodes of the P2P structured botnet is24N+. When N =6, the mean connection hops is 2, it has nice network connectivity, but when N=200 and 2000, the mean connection hops are approximately 50 and 500, respectively, and it larger than our proposed botnet.From the comparison, we can see that our proposed mobile botnet have high connectivity and it can be adjusted easily by adjust the height and scale of mobile botnet. It has high flexibility and scalability.B. SecurityAs we described in section IV, the height of the model is H , 'k is the number of lower tier nodes that communicate with heighter tier node. As is illustrated in Figure 5, the maximum number of nodes at the lowest tieris '()Hk . Then, the maximum number of our proposed botnet model is2''0()()H i Hi N k k −==+∑(1) The larger of the H , the less community load of the key node. But large H will lead to long cummunication path and low connectivity. The relationship between model degree K and node degree i k 〈〉 is12max(,,...,)N K k k k =〈〉〈〉〈〉 (2)'2k k 〈〉≤(3)As we can see form Formula (2) and (3), K increasewith the increase of 'k . During the C&C period, the keynodes have to forward at least 'k messages at one time, but the large 'k means that there is large scale of abnormal data traffic, and that will jeopardize the safetyof key nodes. The relationships of 'k , N , and H are shown in Figure 9. When 'k =3 and the value of H are 2, 3, 4, 5 respectively, the proposed botnet will have better connectivity but have lower capacity. When the value of'k is 5, there is a better leverage between connectivityand N. In this paper, consider the tradeoff between 'k andN , we set our mobile botnet H =5, '5k ≤, so 10k 〈〉≤, 10K ≤, and the scale of the botnet can up to 3281 botnodes.By the tradeoff between k’ and H, we can efficiently control the amount of traffic flow that travels between bot nodes, and the communications between bot nodes will not catch the attention of defender and then protect the mobile botnet, so our mobile botnet can have high security.。