Poster
A Unified, Resilient, and Explainable Adversarial Patch Detector
Vishesh Kumar · Akshay Agarwal
ExHall D Poster #406
Deep Neural Networks (DNNs), backbone architecture in almost′everycomputervisiontask,arev≠̲rab≤→adversarialaacks,partica̲rlyphysicalout-of-distribution(OOD)adversarialpatches.Eξst∈g⊨oftenstrugg≤with∫erpret∈gtheseaacks∈waystˆalignwithhumanvisualperception.Our∝osedAdvPatchXAI∫roducesa≥≠ralized,robust,andexpla∈ab≤defensealgorithmspecificallydesig≠d→defendDℕsaga∈stphysicaladversarialthreats.AdvPatchXAIemploysanovelpatchdecorrelationlosstˆreducesfeatureredundancyandenhancesthedist∈ctive≠ssofpatchrepresentations,enabl∈gbeer≥≠ralizationacrossunseenadversarialscenarios.It≤arnspro→tyπcalparts∈aself-⊇rvisedfashion,enhanc∈g∫erprηbilityandcorrelationwithhumanvision.Themodelutilizesasparsel∈earlayerforclassification,mak∈gthedecision-mak∈gprocessglobally∫erprηb≤throughasetof≤ar≠dpro→typesandlocallyexpla∈ab≤byπnp∮∈gre≤vantpro→typeswith∈anima≥.OurcomprehensiveevaluationshowstˆAdvPatchXAI¬onlyclosesthe`semantic'' gap between latent space and pixel space but also effectively handles unseen adversarial patches even perturbed with unseen corruptions, thereby significantly advancing DNN robustness in practical settings.
Live content is unavailable. Log in and register to view live content