Research map / Adversarial robustness and security / Attacks on text and graphs
Benchmarking Robustness to Adversarial Image Obfuscations
Florian Stimberg, Ayan Chakrabarti, Chun-Ta Lu, Hussein Hazimeh, Otilia Stretcu, Wei Qiao, Yintao Liu, Merve Kaya, Cyrus Rashtchian, Ariel Fuxman, Mehmet Nejat Tek, Sven Gowal
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 |
|---|---|---|---|
| Pretrain on ImageNet-21k (BiT-M vs BiT-S) | BiT models · Obfuscation benchmark | accuracy per obfuscation | Improved |
| Pretrain on ImageNet-21k (ConvNext-Base-21k vs ConvNext-Base) | ConvNext-Base · Obfuscation benchmark | accuracy per obfuscation | Mixed |
| Scale up model parameters or compute | pretrained ImageNet models · Hold-out obfuscations | worst-case accuracy | Improved |
| Add MixUp augmentation | ResNet50 clean training · Hold-out obfuscations | accuracy | Improved |
| Add CutMix augmentation | ResNet50 clean training · Hold-out obfuscations | accuracy | Improved |
| Add AugMix augmentation | ResNet50 clean training · Hold-out obfuscations | accuracy | Mixed |
| Train on PhotoComposition obfuscation | ResNet50 · BackgroundBlurComposition | accuracy: over 30% | Improved |
| Train on BackgroundBlurComposition obfuscation | ResNet50 · PhotoComposition | accuracy: less than 10 % | Improved |
| Add obfuscated images to training | ResNet50 · Clean ImageNet | clean accuracy | No clear effect |
| Use distribution shift algorithms (IRM, DeepCORAL, DANN, JTT) | ResNet50 · Hold-out obfuscations | accuracy | No clear effect |
| Add adversarial training (L_inf eps 4/255) | ResNet50 · Hold-out obfuscations | accuracy | Hurt |
| Pretrain ViT on ImageNet-21k when training with obfuscations | ViT-B16 · Hold-out obfuscations | accuracy | No clear effect |
Closest papers by meaning
- Comparing the Robustness of Modern No-Reference Image- and Video-Quality Metrics to Adversarial Attacks (AAAI 2024 · similarity 0.88)
- Image Safeguarding: Reasoning with Conditional Vision Language Model and Obfuscating Unsafe Content Counterfactually (AAAI 2024 · similarity 0.87)
- OODRobustBench: a Benchmark and Large-Scale Analysis of Adversarial Robustness under Distribution Shift (ICML 2024 · similarity 0.87)
- DiffusionGuard: A Robust Defense Against Malicious Diffusion-based Image Editing (ICLR 2025 · similarity 0.86)
- Robust NAS under adversarial training: benchmark, theory, and beyond (ICLR 2024 · similarity 0.86)
- Guardians of Image Quality: Benchmarking Defenses Against Adversarial Attacks on Image Quality Metrics (ICML 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.
