Research map / Adversarial robustness and security / Attacks on text and graphs
ML-LOO: Detecting Adversarial Examples with Feature Attribution
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 multi-layer LOO attributions rather than the strongest listed detection baseline | Adversarial-example detection · MNIST, CIFAR-10, CIFAR-100; listed attacks | AUC: 0.995 vs 0.990; 0.999 vs 0.970; 0.974 vs 0.951 | Helped |
| Use multi-layer attribution features to detect varied confidence levels | ML-LOO · CIFAR-10, ResNet, C&W-MIX | AUC: 0.840 vs 0.649 | Helped |
| Use multi-layer attribution features to detect varied confidence levels | ML-LOO · CIFAR-10, ResNet, L∞-PGD-MIX | AUC: 0.953 vs 0.813 | Helped |
| Train on C&W examples and transfer detection to other attack types | ML-LOO · CIFAR-10, ResNet, attacks unseen during detector training | AUC: 0.879; 0.992; 0.796; 0.984; 0.994 | Mixed |
| Use LOO feature attribution instead of Integrated Gradients attribution (ablation) | Attribution-based detection · CIFAR-10, ResNet | detection performance | Helped |
| Use IQR rather than STD or MAD to summarize attribution dispersion (ablation) | Attribution-based detection · CIFAR-10, ResNet, C&W, JSMA and L∞-PGD | AUC | Helped |
| Use only output-layer attribution dispersion rather than multi-layer attribution information (ablation) | ML-LOO · Mixed-confidence adversarial examples | detection performance | Hurt |
Closest papers by meaning
- Beating Attackers At Their Own Games: Adversarial Example Detection Using Adversarial Gradient Directions (AAAI 2021 · similarity 0.90)
- On Adversarial Training without Perturbing all Examples (ICLR 2024 · similarity 0.89)
- GAT: Generative Adversarial Training for Adversarial Example Detection and Robust Classification (ICLR 2020 · similarity 0.89)
- Detecting Adversarial Examples Through Image Transformation (AAAI 2018 · similarity 0.88)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · similarity 0.88)
- Wide Two-Layer Networks can Learn from Adversarial Perturbations (NeurIPS 2024 · similarity 0.87)
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.