Research map / Adversarial robustness and security / Backdoor and poisoning
Neural Polarizer: A Lightweight and Effective Backdoor Defense via Purifying Poisoned Features
Mingli Zhu, Shaokui Wei, Hongyuan Zha, Baoyuan Wu
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 the learned neural polarizer with the full NPD objective instead of leaving the backdoored model undefended | PreAct-ResNet18 · CIFAR-10, BadNets-A2O | ASR (%): 93.79 -> 1.26 | Improved |
| Use the learned neural polarizer with the full NPD objective instead of leaving the backdoored model undefended | PreAct-ResNet18 · CIFAR-10, Blended | ASR (%): 97.71 -> 0.41 | Improved |
| Use the learned neural polarizer with the full NPD objective instead of leaving the backdoored model undefended | PreAct-ResNet18 · Tiny ImageNet, 8 attacks (Avg) | ASR (%): 89.64 -> 2.17 | Improved |
| Use known target labels and targeted universal perturbations (NPD-TU) instead of estimated targets and per-sample perturbations (NPD) | NPD · CIFAR-10, 8 attacks (Avg) | ASR (%): 1.62 -> 1.02 | Improved |
| Use known target labels and per-sample targeted perturbations (NPD-TP) instead of estimated targets and per-sample perturbations (NPD) | NPD · CIFAR-10, 8 attacks (Avg) | DER (%): 95.56 -> 96.41 | Improved |
| Use known target labels and targeted perturbations instead of estimated targets | NPD · Tiny ImageNet, 8 attacks (Avg) | DER (%): 90.84 -> 91.46 | Improved |
| Replace targeted perturbations with untargeted universal perturbations (ablation) | NPD · CIFAR-10, LF | ASR (%): 0.21 -> 75.42 | Hurt |
| Replace targeted perturbations with untargeted PGD perturbations (ablation) | NPD · CIFAR-10, LF | ASR (%): 0.21 -> 17.24 | Hurt |
| Add the first boosted-cross-entropy term, which encourages the true label on perturbed samples (ablation) | NPD · CIFAR-10, Blended | ASR (%): 99.63 -> 2.01 | Improved |
| Add the second boosted-cross-entropy term, which suppresses the strongest non-true class (ablation) | NPD · CIFAR-10, LF | ASR (%): 95.90 -> 1.06 | Improved |
| Add the attack-success-rate loss term (ablation) | NPD · CIFAR-10, LF | ASR (%): 95.90 -> 0.80 | Improved |
| Add all three trigger-purification loss terms (ablation) | NPD · CIFAR-10, BadNets-A2O | ASR (%): 1.18 -> 1.26 | Mixed |
Closest papers by meaning
- Uncovering, Explaining, and Mitigating the Superficial Safety of Backdoor Defense (NeurIPS 2024 · similarity 0.85)
- REFINE: Inversion-Free Backdoor Defense via Model Reprogramming (ICLR 2025 · similarity 0.85)
- From Toxic to Trustworthy: Using Self-Distillation and Semi-supervised Methods to Refine Neural Networks (AAAI 2024 · similarity 0.85)
- Unveiling and Mitigating Backdoor Vulnerabilities based on Unlearning Weight Changes and Backdoor Activeness (NeurIPS 2024 · similarity 0.84)
- BAN: Detecting Backdoors Activated by Adversarial Neuron Noise (NeurIPS 2024 · similarity 0.84)
- BaDExpert: Extracting Backdoor Functionality for Accurate Backdoor Input Detection (ICLR 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.
