基于可解释性的面向三维目标跟踪模型的对抗训练方法
打开文本图片集
中图分类号：TP309 文献标志码：A 文章编号：1001-3695（2026）04-012-1061-08
doi：10.19734/j.issn.1001-3695.2025.08.0298
Explainability based adversarial training method for 3D object tracking
Cheng Riranl"，Wang Xupeng1bt，Lei Hangl，Xiao Dian 2，3 ， Yang Qing2，3 （1.a.SchlffoodfeneibofellbiUfl andTechnologofChinaChengdu61，China；2.Sichuan-TibetTechnologyovationCenter（Chengdu）ofNtionalailayCo Ltd.，Chengdu 611432，China；3. China Academy of Railway Sciences Co.，Ltd.，Beijing 10081，China）
Abstract：3Dmodels basedondeepneuralnetworks have made significant progress oncleandatasets.However，their vulnerability to adversarial examples hasledtosecurityrisks inpractical aplications.Toaddresthisissue，this paper proposeda noveladversarialtrainingmethdfor3Dobjecttracking.Firstly，itleveragetheexplainabilitymethodtorevealthecontribution ofeachpointinthe modelinput tothe modelpredictionandstudythechangeinthedecisionof the modelaftertheattck，exploringthecorelationbetweenthesensitivityofapointtoatackanditsimportance.Then，itaddeddversarialexamplescor respondingtocleansamplesduringthetrainingofmodeltogeneratearobust target model.Itencouragedthetargetmodel to align the contributions of corresponding points in clean samples andadversarial samples，ensuring theconsistencyof the model'sdecisions whenfacingadversarial examples.Furthermore，this paper introducedapolicymodel todynamicalladjust the parametersrequiredforgeneratingadversarial examples toensuretheirefectiveness，whichcouldfurtherimprovethero bustness of the model.Extensive experiments on multipledatasets demonstratethat thisapproach enables existing models to achievebeterperformanceagainstadvancedadversarialatackscompared tootherdefense methods.This demonstrates hatthe proposed explainability-basedadversarialtraining method providesafeasibleandeficientsolutionforimproving therobustess of 3Dobject tracking models.
Key Words：3D object tracking；explainability method;defense method;adversarial training
0 引言
近年来，基于深度神经网络的三维目标跟踪得到了迅猛的发展，因其具有广泛的应用前景，例如自动驾驶、机器人科学、视频监控等[1～5]。（剩余22060字）