Research map / Adversarial robustness and security / Certified and empirical robustness
Understanding Sensitivity of Differential Attention through the Lens of Adversarial Robustness
Tsubasa Takahashi, Shojiro Yamabe, Futa Kai Waseda, Kento Sasaki
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace standard attention with differential attention | ViT (1 layer) · CIFAR-10/100 adversarial examples and patches | ASR | Hurt |
| replace standard attention with differential attention | ViT (depth 4) · Tiny ImageNet PGD eps=1/255 | ASR: 0.2309 -> 0.2822 | Hurt |
| replace standard attention with differential attention | ViT (depth 4) · Tiny ImageNet PGD eps=4/255 | ASR: 0.6490 -> 0.6318 | No clear effect |
| use DiffCLIP instead of CLIP | CLIP ViT-B/16 · COCO | ASR | Hurt |
| use DiffCLIP instead of CLIP | CLIP ViT-B/16 · ImageNet-1k | ASR | Mixed |
| raise lambda_init from 0.5 to 0.8 (ablation) | DiffViT · CIFAR-10 PGD eps=1/255 | ASR: 0.4074 -> 0.8498 | Hurt |
| raise lambda_init from 0.8 to 0.95 (ablation) | DiffViT · CIFAR-10 PGD eps=1/255 | ASR: 0.8498 -> 0.4164 | Improved |
| increase DA depth (stack layers) (ablation) | DiffViT · PGD/AutoAttack eps=1/255 | ASR | Improved |
| increase DA depth (stack layers) (ablation) | DiffViT · PGD/AutoAttack eps=4/255 | ASR | No clear effect |
| increase DA depth (ablation) | DiffViT · Mean local Lipschitz estimate | local Lipschitz | Hurt |
| add adversarial training eps=0.5/255 (ablation) | DiffViT · CIFAR-10 PGD eps=0.5/255 | ASR: 0.6367 -> 0.4440 | Improved |
| add adversarial training eps=1/255 (ablation) | DiffViT · CIFAR-10 clean accuracy | accuracy: 0.8700 -> 0.7384 | Hurt |
Closest papers by meaning
- Stage-wise Attention-Guided Region Sequencing for Adversarial Attacks on Large Vision-Language Models (NeurIPS 2026 · similarity 0.85)
- Perturbing Attention Gives You More Bang for the Buck: Subtle Imaging Perturbations That Efficiently Fool Customized Diffusion Models (CVPR 2024 · similarity 0.85)
- The Implicit Bias of Gradient Descent toward Collaboration between Layers: A Dynamic Analysis of Multilayer Perceptions (NeurIPS 2024 · similarity 0.85)
- Understanding Differential Transformer Unchains Pretrained Self-Attentions (NeurIPS 2025 · similarity 0.85)
- Text-Guided Attention is All You Need for Zero-Shot Robustness in Vision-Language Models (NeurIPS 2024 · similarity 0.85)
- Mind the Gap: a Spectral Analysis of Rank Collapse and Signal Propagation in Attention Layers (ICML 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
