Research map / Adversarial robustness and security / Backdoor and poisoning
Spectral Signatures in Backdoor Attacks
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 |
|---|---|---|---|
| Apply spectral outlier removal to learned representations and retrain | ResNet backdoor defense · CIFAR-10 poisoned test set, target bird, 5% poison | poisoned accuracy: 74.20% -> 2.00% | Helped |
| Apply spectral outlier removal to learned representations and retrain | ResNet backdoor defense · CIFAR-10 poisoned test set, target cat, 10% poison | poisoned accuracy: 92.00% -> 0.00% | Helped |
| Apply spectral outlier removal to learned representations and retrain | ResNet backdoor defense · CIFAR-10 clean test set, target cat, 10% poison | natural accuracy: 92.39% -> 92.44% | No clear effect |
| Apply spectral outlier removal to learned representations and retrain | ResNet backdoor defense · CIFAR-10 poisoned test set, target cat, 10% poison, second cat experiment | poisoned accuracy: 99.10% -> 8.20% | No clear effect |
| Apply spectral outlier removal to learned representations and retrain | ResNet backdoor defense · CIFAR-10 poisoned test set, target horse, 5% poison | poisoned accuracy: 99.80% -> 1.00% | Helped |
| Use the top singular-vector correlation score instead of representation norms or correlation with a random vector (ablation) | Spectral signatures · CIFAR-10 clean and poisoned training examples | poison detection | Helped |
| Apply robust-statistics outlier removal to raw inputs instead of learned representations (ablation) | Spectral signatures · CIFAR-10 backdoor training data | poison detection | Hurt |
| Use learned representations from later network layers rather than the input (ablation) | Spectral signatures · CIFAR-10 backdoor training data | poison detection | Helped |
| Combine cats and dogs into a heterogeneous pets class (ablation) | Spectral signatures · CIFAR-10 combined-label backdoor experiments | poisoned accuracy after removal | No clear effect |
| Use VGG instead of ResNet as the feature model (ablation) | Spectral signatures · CIFAR-10 backdoor detection | poisoned accuracy after removal | No clear effect |
Closest papers by meaning
- Demystifying Poisoning Backdoor Attacks from a Statistical Perspective (ICLR 2024 · similarity 0.82)
- Universal Backdoor Attacks (ICLR 2024 · similarity 0.81)
- Activation Gradient based Poisoned Sample Detection Against Backdoor Attacks (ICLR 2025 · similarity 0.80)
- Protecting against simultaneous data poisoning attacks (ICLR 2025 · similarity 0.79)
- Adversarial-Inspired Backdoor Defense via Bridging Backdoor and Adversarial Attacks (AAAI 2025 · similarity 0.79)
- Mitigating Backdoor Attack by Injecting Proactive Defensive Backdoor (NeurIPS 2024 · similarity 0.78)
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.