Matches in SemOpenAlex for { <https://semopenalex.org/work/W2982584252> ?p ?o ?g. }
- W2982584252 abstract "We propose a simple change to existing neural network structures for better defending against gradient-based adversarial attacks. Instead of using popular activation functions (such as ReLU), we advocate the use of k-Winners-Take-All (k-WTA) activation, a C0 discontinuous function that purposely invalidates the neural network model's gradient at densely distributed input data points. The proposed k-WTA activation can be readily used in nearly all existing networks and training methods with no significant overhead. Our proposal is theoretically rationalized. We analyze why the discontinuities in k-WTA networks can largely prevent gradient-based search of adversarial examples and why they at the same time remain innocuous to the network training. This understanding is also empirically backed. We test k-WTA activation on various network structures optimized by a training method, be it adversarial training or not. In all cases, the robustness of k-WTA networks outperforms that of traditional networks under white-box attacks." @default.
- W2982584252 created "2019-11-08" @default.
- W2982584252 creator A5025499043 @default.
- W2982584252 creator A5062439923 @default.
- W2982584252 creator A5068853170 @default.
- W2982584252 date "2019-05-25" @default.
- W2982584252 modified "2023-09-27" @default.
- W2982584252 title "Enhancing Adversarial Defense by k-Winners-Take-All" @default.
- W2982584252 cites W1682403713 @default.
- W2982584252 cites W1686810756 @default.
- W2982584252 cites W187321743 @default.
- W2982584252 cites W1938785624 @default.
- W2982584252 cites W2099049980 @default.
- W2982584252 cites W2099896495 @default.
- W2982584252 cites W2108598243 @default.
- W2982584252 cites W2112796928 @default.
- W2982584252 cites W2125908420 @default.
- W2982584252 cites W2141700268 @default.
- W2982584252 cites W2149194912 @default.
- W2982584252 cites W2151298633 @default.
- W2982584252 cites W2161388792 @default.
- W2982584252 cites W2167148127 @default.
- W2982584252 cites W2194775991 @default.
- W2982584252 cites W2243397390 @default.
- W2982584252 cites W2335728318 @default.
- W2982584252 cites W2342045095 @default.
- W2982584252 cites W2401231614 @default.
- W2982584252 cites W2535873859 @default.
- W2982584252 cites W2561498661 @default.
- W2982584252 cites W2603766943 @default.
- W2982584252 cites W2735607295 @default.
- W2982584252 cites W2736899637 @default.
- W2982584252 cites W2752885492 @default.
- W2982584252 cites W2774644650 @default.
- W2982584252 cites W2786118190 @default.
- W2982584252 cites W2886067286 @default.
- W2982584252 cites W2887603965 @default.
- W2982584252 cites W2894604724 @default.
- W2982584252 cites W2897355816 @default.
- W2982584252 cites W2899748887 @default.
- W2982584252 cites W2903785932 @default.
- W2982584252 cites W2913266441 @default.
- W2982584252 cites W2932048622 @default.
- W2982584252 cites W2938470389 @default.
- W2982584252 cites W2947657997 @default.
- W2982584252 cites W2948046738 @default.
- W2982584252 cites W2950048339 @default.
- W2982584252 cites W2962759300 @default.
- W2982584252 cites W2963001136 @default.
- W2982584252 cites W2963070423 @default.
- W2982584252 cites W2963143631 @default.
- W2982584252 cites W2963158386 @default.
- W2982584252 cites W2963207607 @default.
- W2982584252 cites W2963249138 @default.
- W2982584252 cites W2963389226 @default.
- W2982584252 cites W2963440492 @default.
- W2982584252 cites W2963446712 @default.
- W2982584252 cites W2963570359 @default.
- W2982584252 cites W2963744840 @default.
- W2982584252 cites W2963857521 @default.
- W2982584252 cites W2963920068 @default.
- W2982584252 cites W2963982496 @default.
- W2982584252 cites W2964014389 @default.
- W2982584252 cites W2964153729 @default.
- W2982584252 cites W2964197269 @default.
- W2982584252 cites W2964224652 @default.
- W2982584252 cites W2964253222 @default.
- W2982584252 cites W2970317235 @default.
- W2982584252 cites W2991290085 @default.
- W2982584252 cites W3103557498 @default.
- W2982584252 hasPublicationYear "2019" @default.
- W2982584252 type Work @default.
- W2982584252 sameAs 2982584252 @default.
- W2982584252 citedByCount "5" @default.
- W2982584252 countsByYear W29825842522019 @default.
- W2982584252 countsByYear W29825842522020 @default.
- W2982584252 countsByYear W29825842522021 @default.
- W2982584252 crossrefType "posted-content" @default.
- W2982584252 hasAuthorship W2982584252A5025499043 @default.
- W2982584252 hasAuthorship W2982584252A5062439923 @default.
- W2982584252 hasAuthorship W2982584252A5068853170 @default.
- W2982584252 hasConcept C104317684 @default.
- W2982584252 hasConcept C111919701 @default.
- W2982584252 hasConcept C11413529 @default.
- W2982584252 hasConcept C134306372 @default.
- W2982584252 hasConcept C14036430 @default.
- W2982584252 hasConcept C154945302 @default.
- W2982584252 hasConcept C15627037 @default.
- W2982584252 hasConcept C185592680 @default.
- W2982584252 hasConcept C2779960059 @default.
- W2982584252 hasConcept C2984842247 @default.
- W2982584252 hasConcept C33923547 @default.
- W2982584252 hasConcept C37736160 @default.
- W2982584252 hasConcept C38365724 @default.
- W2982584252 hasConcept C41008148 @default.
- W2982584252 hasConcept C50644808 @default.
- W2982584252 hasConcept C55493867 @default.
- W2982584252 hasConcept C63479239 @default.
- W2982584252 hasConcept C78458016 @default.
- W2982584252 hasConcept C86803240 @default.