notesum.ai
Published at October 30Backdoor Attack Against Vision Transformers via Attention Gradient-Based Image Erosion
cs.CV
cs.AI
Released Date: October 30, 2024
Authors: Ji Guo1, Hongwei Li2, Wenbo Jiang3, Guoming Lu1
Aff.: 1Laboratory Of Intelligent Collaborative Computing, University of Electronic Science and Technology of China, China; 2School of Computer Science and Engineering, University of Electronic Science and Technology of China, China; 3School of Computer Science and Engineering, University of Electronic Science and Technology of China, China (Corresponding Author)

| Datasets | ViT-b | Deit-t | Deit-s | ||||||
|---|---|---|---|---|---|---|---|---|---|
| ACC | ASR | CDA | ACC | ASR | CDA | ACC | ASR | CDA | |
| CIFAR-10 | 98.64% | 99.85% | 98.63% | 94.67% | 99.83% | 94.66% | 96.74% | 99.78% | 96.46% |
| GBSTR | 96.76% | 97.82% | 94.20% | 95.14% | 97.85% | 94.59% | 95.46% | 96.16% | 93.64% |
| ImageNette | 99.59% | 95.13% | 98.29% | 96.80% | 95.27% | 96.24% | 98.11% | 97.02% | 98.06% |