Research map / Adversarial robustness and security / Attacks on text and graphs
Sample Efficient Detection and Classification of Adversarial Attacks via Self-Supervised Embeddings
Mazda Moayeri, Soheil Feizi
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 SimCLR self-supervised embeddings instead of supervised ResNet50 embeddings | SimCat · ImageNet 8-attack classification (5 samples per attack) | accuracy (%): +7.7 | Improved |
| use SimCLR embeddings instead of supervised embeddings | SimCat · ImageNet detection (5 samples per attack) | accuracy (%): 68.5 (+6.4) | Improved |
| use SimCLR embeddings instead of supervised embeddings | SimCat · SVHN PGD-l2 vs PGD-linf classification (25 samples) | accuracy (%): 77.1 (+21.0) | Improved |
| increase training samples per attack from 5 to 100 | SimCat · ImageNet 8-attack classification | accuracy (%): 27.1 -> 58.1 | Improved |
| increase training samples per attack from 2 to 50 | SimCat · SVHN PGD-linf detection | accuracy (%): 77.3 -> 94.2 | Improved |
| train on union of attacks with 5 samples each instead of one attack with 100 samples | SimCat · ImageNet unseen-attack detection | accuracy (%): 64.7 -> 66.5 | Improved |
| train on union of attacks with 20 samples each instead of 5 | SimCat · ImageNet unseen-attack detection | accuracy (%): 66.5 -> 71.1 | Improved |
| use ensemble of separate BP and CP detectors instead of one joint detector (ablation) | SimCat · STL10 poison defense (avg PSR) | poison success rate (%): 13.9 -> 9.7 | Improved |
| add SimCat poison defense | STL10 linear transfer model · STL10 poison defense (avg PSR) | poison success rate (%): 21.8 -> 9.7 | Improved |
| add SimCat ensemble poison defense | STL10 linear transfer model · STL10 clean accuracy | accuracy (%): 86 -> 85 | Hurt |
| exclude FC poisons from general detection (ablation) | SimCat · STL10 general poison detection | accuracy (%): 64.5 -> 70.5 | Improved |
| add adaptive adversarial training (ablation) | SimCat · ImageNet detection under adaptive PGD-l2 | robustness (%): 39.25 -> 67.95 | Improved |
Closest papers by meaning
- Self-Supervised Representation Learning for Adversarial Attack Detection (ECCV 2024 · similarity 0.87)
- AttackBench: Evaluating Gradient-based Attacks for Adversarial Examples (AAAI 2025 · similarity 0.85)
- ASTrA: Adversarial Self-supervised Training with Adaptive-Attacks (ICLR 2025 · similarity 0.85)
- Extending Adversarial Attacks to Produce Adversarial Class Probability Distributions (JMLR 2023 · similarity 0.85)
- Alleviating Adversarial Attacks on Variational Autoencoders with MCMC (NeurIPS 2022 · similarity 0.85)
- Taxonomy Driven Fast Adversarial Training (AAAI 2024 · 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.
