Research map / Adversarial robustness and security / Certified and empirical robustness
DiffAttack: Evasion Attacks Against Diffusion-Based Adversarial Purification
Mintong Kang, Dawn Song, Bo Li
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 |
|---|---|---|---|
| Use DiffAttack instead of AdjAttack | score-based purification, WideResNet-28-10 · CIFAR-10 linf 8/255 | robust accuracy (%): 70.64 -> 46.88 | Improved |
| Use DiffAttack instead of AdjAttack | score-based purification, WideResNet-70-16 · CIFAR-10 linf 8/255 | robust accuracy (%): 71.29 -> 45.31 | Improved |
| Use DiffAttack instead of AdjAttack | score-based purification, WideResNet-28-10 · CIFAR-10 linf 4/255 | robust accuracy (%): 82.81 -> 71.88 | Improved |
| Use DiffAttack instead of AdjAttack | score-based purification, WideResNet-28-10 · CIFAR-10 l2 0.5 | robust accuracy (%): 78.58 -> 64.06 | Improved |
| Use DiffAttack instead of AdjAttack | score-based purification, ResNet-50 · ImageNet linf 4/255 | robust accuracy (%): 40.93 -> 28.13 | Improved |
| Use DiffAttack instead of Diff-BPDA | DDPM-based purification, WideResNet-28-10 · CIFAR-10 linf 8/255 | robust accuracy (%): 75.00 -> 54.69 | Improved |
| Use DiffAttack instead of Diff-BPDA | DDPM-based purification, WideResNet-70-16 · CIFAR-10 l2 0.5 | robust accuracy (%): 81.25 -> 71.88 | Improved |
| Use DiffAttack instead of SPSA | score-based purification · CIFAR-10 linf 8/255 | robust accuracy (%): 83.37 -> 46.88 | Improved |
| Use DiffAttack instead of Square Attack | DDPM-based purification · CIFAR-10 linf 8/255 | robust accuracy (%): 81.68 -> 54.69 | Improved |
| Use DiffAttack instead of Joint Attack (Full) | DDPM-based purification · CIFAR-10 linf 8/255 | robust accuracy (%): 76.26 -> 54.69 | Improved |
| Use segment-wise forwarding-backwarding instead of full computational graph (ablation) | gradient backpropagation through diffusion · CIFAR-10 WideResNet-28-10, batch 16 | memory cost | Improved |
| Increase diffusion length T of the purifier (ablation) | diffusion-based purification · CIFAR-10 WideResNet-28-10 | clean accuracy | Hurt |
Later papers that compared with this method
11 later accepted papers report DiffAttack as a baseline in their results tables; 8 of them report a better number on most of the results they share with it.
- Beyond Single-Point Perturbation: A Hierarchical, Manifold-Aware Approach to Diffusion Attacks (AAAI 2026 · better on 13 of 14 shared results)
- Concept-based Adversarial Attack: a Probabilistic Perspective (ICLR 2026 · better on 8 of 9 shared results)
- AdvFM: Lookahead Flow-Matching Velocity-Field Attacks for Imperceptible and Transferable Adversarial Examples (CVPR 2026 · better on 5 of 6 shared results)
- DASH: A Meta-Attack Framework for Synthesizing Effective and Stealthy Adversarial Examples (CVPR 2026 · better on 6 of 6 shared results)
- ObjectAdv: Object-Level Unrestricted Adversarial Attacks via Diffusion Models (AAAI 2026 · better on 4 of 4 shared results)
- Generative Adversarial Perturbations with Cross-paradigm Transferability on Localized Crowd Counting (CVPR 2026 · better on 1 of 3 shared results)
- NatADiff: Adversarial Boundary Guidance for Natural Adversarial Diffusion (ICLR 2026 · better on 1 of 2 shared results)
- Untraceable DeepFakes via Traceable Fingerprint Elimination (ICLR 2026 · better on 1 of 1 shared result)
3 more not listed.
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Square Attack — Square Attack: a query-efficient black-box adversarial attack via random search (ECCV 2020) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- DiffHammer: Rethinking the Robustness of Diffusion-Based Adversarial Purification (NeurIPS 2024 · similarity 0.93)
- DiffBreak: Is Diffusion-Based Purification Robust? (NeurIPS 2025 · similarity 0.91)
- LoRID: Low-Rank Iterative Diffusion for Adversarial Purification (AAAI 2025 · similarity 0.91)
- Towards Understanding the Robustness of Diffusion-Based Purification: A Stochastic Perspective (ICLR 2025 · similarity 0.91)
- Divide and Conquer: Heterogeneous Noise Integration for Diffusion-based Adversarial Purification (CVPR 2025 · similarity 0.90)
- Instant Adversarial Purification with Adversarial Consistency Distillation (CVPR 2025 · similarity 0.90)
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.
