Research map / Adversarial robustness and security / Certified and empirical robustness
DRIFT: Divergent Response in Filtered Transformations for Robust Adversarial Defense
Amira Guesmi, Muhammad Shafique
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 |
|---|---|---|---|
| add Jacobian separation loss (ablation) | CE + adversarial training · ImageNet adaptive PGD (ResNet-v2) | robust accuracy: 3.70 -> 39.80 | Improved |
| add logit-VJP separation loss (ablation) | CE + adversarial training · ImageNet adaptive PGD (ResNet-v2) | robust accuracy: 3.70 -> 47.61 | Improved |
| combine JS and LVJP losses (ablation) | CE + LVJP + adv · ImageNet adaptive PGD (ResNet-v2) | robust accuracy: 47.61 -> 53.78 | Improved |
| combine JS and LVJP losses (ablation) | CE + LVJP + adv · ImageNet adaptive PGD (ViT-B/16) | robust accuracy: 49.73 -> 56.74 | Improved |
| separation losses vs clean accuracy (ablation) | DRIFT · ImageNet non-adaptive | clean accuracy: 75.66 -> 76.19 (ResNet-v2) | No clear effect |
| use residual two-conv filter instead of single conv (ablation) | DRIFT filter · ImageNet adaptive PGD (ViT-B/16) | robust accuracy: 42.5 -> 56.7 | Improved |
| use residual two-conv filter instead of deeper four-conv filter (ablation) | DRIFT filter · ImageNet adaptive PGD (ViT-B/16) | robust accuracy: 47.9 -> 56.7 | Improved |
| allow attacker BPDA in addition to EOT | DRIFT · ImageNet adaptive PGD (ResNet-v2) | robust accuracy: 53.78 -> 60.19 | Mixed |
| learned filter ensemble vs diffusion purification | DiffPure · ImageNet AutoAttack non-adaptive (ResNet-v2) | robust accuracy: 67.01 -> 74.30 | Improved |
| increase EOT samples K from 5 to 20 | DRIFT · ImageNet PGD l_inf eps 4/255 (ViT-B/16) | robust accuracy: 57.62 -> 38.57 | Hurt |
| replace diffusion purification with learned filters | DiffPure · ResNet-50 latency | latency (s): 17.07 -> 0.0004 | Improved |
Closest papers by meaning
- Consensus-Robust Transfer Attacks via Parameter and Representation Perturbations (NeurIPS 2025 · similarity 0.89)
- Demystifying the Adversarial Robustness of Random Transformation Defenses (ICML 2022 · similarity 0.88)
- Enhancing Adversarial Transferability with Checkpoints of a Single Model's Training (CVPR 2025 · similarity 0.86)
- Bridging Symmetry and Robustness: On the Role of Equivariance in Enhancing Adversarial Robustness (NeurIPS 2025 · similarity 0.86)
- Towards Understanding the Robustness of Diffusion-Based Purification: A Stochastic Perspective (ICLR 2025 · similarity 0.86)
- Rethinking Invariance Regularization in Adversarial Training to Improve Robustness-Accuracy Trade-off (ICLR 2025 · similarity 0.86)
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.
