Research map / Adversarial robustness and security / Backdoor and poisoning
SPECTRE: defending against backdoor attacks using robust statistics
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 SPECTRE instead of PCA defense | CIFAR-10 model after 1-way pixel poisoning · 1-way pixel attack, 250 poisons | backdoor test accuracy: 0.880 -> 0.001 | Helped |
| apply SPECTRE instead of PCA defense | CIFAR-10 model after 2-way pixel poisoning · 2-way pixel attack, 125 poisons | backdoor test accuracy: 0.139 -> 0.000 | Helped |
| apply SPECTRE instead of PCA defense | CIFAR-10 model after 1-way periodic poisoning · 1-way periodic attack, 250 poisons | backdoor test accuracy: 0.968 -> 0.001 | Helped |
| apply SPECTRE instead of PCA defense | CIFAR-10 model after label-consistent ℓ₂ poisoning · label-consistent ℓ₂ attack, 125 poisons | poisoned examples removed: 1 -> 125 | Helped |
| apply SPECTRE instead of PCA defense | CIFAR-10 model after label-consistent GAN poisoning · label-consistent GAN attack, 125 poisons | poisoned examples removed: 28 -> 125 | Helped |
| add robust covariance estimation and whitening (ablation) | adaptive dimension reduction + QUE · pixel attack, m=1, 125 poisons | poisoned examples removed: 0 -> 124 | Helped |
| add robust covariance estimation (ablation) | adaptive dimension reduction + whitening + QUE using sample covariance · pixel attack, m=1, 125 poisons | poisoned examples removed: 51 -> 124 | Helped |
| add robust covariance estimation and whitening (ablation) | adaptive dimension reduction + QUE · periodic attack, m=1, 125 poisons | poisoned examples removed: 0 -> 124 | Helped |
| add robust covariance estimation (ablation) | adaptive dimension reduction + whitening + QUE using sample covariance · periodic attack, m=1, 125 poisons | poisoned examples removed: 67 -> 124 | Helped |
| use QUE scoring with α=4 instead of squared norm scoring (ablation) | whitened representations in SPECTRE · 1-way pixel attack, ε=0.1 | remaining poisoned samples: 3 -> 0 | Helped |
| use QUE scoring with α=4 instead of top-principal-direction scoring (ablation) | whitened representations in SPECTRE · 1-way pixel attack, ε=0.1 | remaining poisoned samples: 118 -> 0 | Helped |
| use a fixed dimension k instead of adaptively selecting k (ablation) | SPECTRE · multiple attacks | minimum fraction of poisons removed: 87.1 -> 4.0 | Hurt |
Closest papers by meaning
- Backdoor Secrets Unveiled: Identifying Backdoor Data with Optimized Scaled Prediction Consistency (ICLR 2024 · similarity 0.86)
- Protecting against simultaneous data poisoning attacks (ICLR 2025 · similarity 0.85)
- RPP: A Certified Poisoned-Sample Detection Framework for Backdoor Attacks under Dataset Imbalance (NeurIPS 2026 · similarity 0.85)
- Mitigating Backdoor Attack by Injecting Proactive Defensive Backdoor (NeurIPS 2024 · similarity 0.85)
- The Good, the Bad and the Ugly: Meta-Analysis of Watermarks, Transferable Attacks and Adversarial Defenses (NeurIPS 2025 · similarity 0.84)
- Defense against Model Extraction Attack by Bayesian Active Watermarking (ICML 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.