Research map / Adversarial robustness and security / Certified and empirical robustness
Adversarial Robustness of Discriminative Self-Supervised Learning in Vision
Ömer Veysel Çağatan, Ömer Faruk Tal, M. Emre Gursoy
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 self-supervised pretraining instead of supervised pretraining (frozen backbone) | ResNet50 · ImageNet adversarial robustness | accuracy under attack: 34.5 (supervised) vs 40.5 (MoCoV3), 32.6 (SwAV), 32.1 (SimCLR) | Mixed |
| add non-linear projector and stronger augmentation (MoCoV1 to MoCoV2) (ablation) | MoCo · ImageNet overall adversarial | accuracy drop (%): 71 -> 52 | Improved |
| add prediction head and larger batch size (MoCoV2 to MoCoV3) (ablation) | MoCo · ImageNet overall adversarial | accuracy drop (%): 52 -> 46 | Improved |
| use ViT backbone instead of ResNet (ablation) | DINO · ImageNet overall adversarial | accuracy drop (%): 55% -> 42% drop | Improved |
| use ViT backbone instead of ResNet (ablation) | MoCoV3 · ImageNet overall adversarial | accuracy drop (%): 46% -> 61% drop | Hurt |
| train longer (more epochs) (ablation) | SwAV and MoCoV3 · ImageNet IAA | adversarial robustness: about 1% marginal improvement | No clear effect |
| train longer (beyond 100 epochs) (ablation) | SwAV and MoCoV3 · ImageNet UAP | adversarial accuracy (%): SwAV 33 -> 40; MoCoV3 41 -> 48 | Improved |
| use full augmentation set instead of limited augmentation (ablation) | BYOL · ImageNet UAP | accuracy drop (%): BYOL-NC 46, BYOL-CC 48 -> standard 39 | Improved |
| vary batch size (128 vs 512 vs standard) (ablation) | BYOL · ImageNet IAA/UAP | accuracy drop (%): IAA 57, 56, 56; UAP 41, 43, 39 | No clear effect |
| fine-tune instead of linear probe | SSL and supervised ResNet50 · Transfer learning (adversarial drop) | accuracy drop (%): IAA 60 -> 49; UAP 49 -> 43 | Improved |
| fine-tune instead of linear probe | ImageNet-to-downstream IAA robustness correlation · Transfer learning | Spearman correlation of drops: 0.79 -> 0.12 | Hurt |
| fine-tune instead of linear probe | SwAV · CIFAR-10 FGSM | accuracy drop: 79% -> 17% | Improved |
Closest papers by meaning
- Contrastive Self-Supervised Learning Leads to Higher Adversarial Susceptibility (AAAI 2023 · similarity 0.91)
- 3SAT: A Simple Self-Supervised Adversarial Training Framework (AAAI 2025 · similarity 0.88)
- Robust SAM: On the Adversarial Robustness of Vision Foundation Models (AAAI 2025 · similarity 0.88)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · similarity 0.88)
- Adversarial Robustness Limits via Scaling-Law and Human-Alignment Studies (ICML 2024 · similarity 0.88)
- Explaining the role of Intrinsic Dimensionality in Adversarial Training (ICML 2025 · similarity 0.88)
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.
