Research map / Adversarial robustness and security / Backdoor and poisoning
CLEAR: Clean-Up Sample-Targeted Backdoor in Neural Networks
Liuwan Zhu, Rui Ning, Chunsheng Xin, Chonggang Wang, Hongyi 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 |
|---|---|---|---|
| patch with generated pocket samples instead of clean samples only (ablation) | CLEAR · Bullseye multi-target, transfer learning (CIFAR10/Multi-View Car) | attack success rate: 90.0% -> 8.8% | Improved |
| patch with generated pocket samples instead of clean samples only (ablation) | CLEAR · Label Flipping, transfer learning (CIFAR10) | attack success rate: 45.0% -> 10.0% | Improved |
| patch with generated pocket samples instead of clean samples only (ablation) | CLEAR · Label Flipping, end-to-end (CIFAR10) | attack success rate: 100% -> 40.0% | Improved |
| patch with generated pocket samples instead of clean samples only (ablation) | CLEAR · Bullseye Polytope multi-target, end-to-end (CIFAR10/Multi-View Car) | attack success rate: 92.8% -> 15.7% | Improved |
| patch with generated pocket samples instead of clean samples only (ablation) | CLEAR · Label Flipping, transfer learning (ImageNet) | attack success rate: 95% -> 0% | Improved |
| patch with generated pocket samples instead of clean samples only (ablation) | CLEAR · Bullseye Polytope single-target, end-to-end (ImageNet) | attack success rate: 95% -> 20% | Improved |
| patch with clean samples instead of generated samples on Feature Collision (ablation) | CLEAR · Feature Collision, transfer learning (CIFAR10) | attack success rate: 0% -> 5.0% | No clear effect |
| patch with generated pocket samples (ablation) | CLEAR · CIFAR10 clean test accuracy, Label Flipping transfer | test accuracy: 94.7% -> 94.9% | No clear effect |
| use CLEAR patching instead of fine-pruning | Fine-pruning · Convex Polytope, transfer learning (CIFAR10) | attack success rate: 51.25% -> 3.8% | Improved |
| fine-tune the entire model (end-to-end) instead of only the last layer when planting the backdoor | CLEAR · Bullseye Polytope multi-target detection | detection success rate: 93.8% -> 87.1% | Hurt |
| fine-tune the entire model (end-to-end) instead of only the last layer on ImageNet | CLEAR · Label Flipping detection (ImageNet) | detection success rate: 95% -> 60% | Hurt |
Closest papers by meaning
- Wicked Oddities: Selectively Poisoning for Effective Clean-Label Backdoor Attacks (ICLR 2025 · similarity 0.90)
- Activation Gradient based Poisoned Sample Detection Against Backdoor Attacks (ICLR 2025 · similarity 0.89)
- Backdoor Defense via Test-Time Detecting and Repairing (CVPR 2024 · similarity 0.89)
- A Theoretical Analysis of Backdoor Poisoning Attacks in Convolutional Neural Networks (ICML 2024 · similarity 0.88)
- Demystifying Poisoning Backdoor Attacks from a Statistical Perspective (ICLR 2024 · similarity 0.88)
- Adversarial-Inspired Backdoor Defense via Bridging Backdoor and Adversarial Attacks (AAAI 2025 · similarity 0.88)
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.
