Research map / Adversarial robustness and security / Attacks on text and graphs
Self-Supervised Representation Learning for Adversarial Attack Detection
Yi Li, Plamen Angelov, Neeraj Suri
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 |
|---|---|---|---|
| replace supervised and pre-trained detectors with self-supervised PAA-ResNet-L | adversarial attack detection · ImageNet-R | detection rate on attacked images (%): 82.8 -> 86.8 (vs DINOV2) | Improved |
| PAA-ResNet-S vs DINOV2 on cost | detector · ImageNet-R | latency (ms): 572.0 -> 6.2 | Improved |
| add prototype-wise contrastive estimation L_PCE on top of L_PM (ablation) | PAA-ResNet-L · ImageNet-R | attacked detection rate (%): 68.9 -> 82.8 | Improved |
| add instance-wise contrastive loss L_ICL with discrimination bank (ablation) | L_PM + L_PCE · ImageNet-R | attacked detection rate (%): 82.8 -> 86.8 | Improved |
| replace L_PM with L_ICL in combination with L_PCE (ablation) | L_PCE + L_ICL · ImageNet-R | attacked detection rate (%): 82.8 -> 82.0 | Hurt |
| use L_PCE instead of prototype loss L_S2Z (ablation) | L_PM-based training · ImageNet-R | attacked detection rate (%): 68.6 -> 72.4 | Improved |
| add data augmentation alone to baseline (ablation) | PAA-ResNet-L · ImageNet-R | attacked detection rate (%): 52.3 -> 50.1 | Hurt |
| use PAA blocks instead of axial attention (ablation) | ResNet-50, HRNet, Xception · ImageNet-R | detection accuracy: 2.1% to 2.9% over axial-attention | Improved |
| vary temperature tau from 0.03 to 0.2 | L_PM · ImageNet-R | detection accuracy | No clear effect |
| weight L_PCE and L_ICL tenfold relative to L_PM | total loss · ImageNet-R | detection accuracy | No clear effect |
| test on unseen dataset instead of seen dataset | all detectors · cross-dataset (Table 6) | detection rate | Hurt |
Closest papers by meaning
- Sample Efficient Detection and Classification of Adversarial Attacks via Self-Supervised Embeddings (ICCV 2021 · similarity 0.87)
- ASTrA: Adversarial Self-supervised Training with Adaptive-Attacks (ICLR 2025 · similarity 0.87)
- A Self-supervised Approach for Adversarial Robustness (CVPR 2020 · similarity 0.86)
- Be Your Own Neighborhood: Detecting Adversarial Examples by the Neighborhood Relations Built on Self-Supervised Learning (ICML 2024 · similarity 0.85)
- Effective Targeted Attacks for Adversarial Self-Supervised Learning (NeurIPS 2023 · similarity 0.85)
- Self-Supervised Adversarial Training via Diverse Augmented Queries and Self-Supervised Double Perturbation (NeurIPS 2024 · similarity 0.83)
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.
