Research map / Adversarial robustness and security / Backdoor and poisoning
One-Pixel Signature: Characterizing CNN Models for Backdoor Detection
Shanjiaoyang Huang, Weiqi Peng, Zhiwei Jia, Zhuowen Tu
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 one-pixel signature detector instead of Neural Cleanse | backdoor detector · MNIST | detection rate: 60.53% -> 95.72% | Improved |
| use one-pixel signature detector instead of Neural Cleanse | backdoor detector · fashion-MNIST | detection rate: 63.16% -> 94.59% | Improved |
| use one-pixel signature detector instead of Neural Cleanse | backdoor detector · GTSRB | detection rate: 60.71% -> 88.47% | Improved |
| use one-pixel signature detector instead of Neural Cleanse | backdoor detector · ImageNet (10-class subset) | detection rate: 54.42% -> 88.69% | Improved |
| use one-pixel signature detector instead of ABS | backdoor detector · GTSRB | detection rate: 66.33% -> 88.47% | Improved |
| test on unseen architecture without training on it | One-Pixel Signature detector · MNIST | detection rate: 95.72% -> 76.50 ± 12.07 (lowest, ResNet unseen) | Hurt |
| train on 175+175 models instead of 300 pairs (ablation) | One-Pixel Signature detector · GTSRB | detection rate: 88.47% -> 87.31% | No clear effect |
| add per-channel samples as training examples (ablation) | One-Pixel Signature detector · GTSRB | training pairs needed: 300 pairs -> 15 pairs for same rate | Improved |
| add per-channel samples as training examples (ablation) | One-Pixel Signature detector · GTSRB | detection rate: 97% with 100 pairs | Improved |
| use one-pixel signature instead of Neural Cleanse/ABS on all-to-all attack | backdoor detector · all-to-all attack | detection success rate: around 50% -> above 95% | Improved |
| use all-to-one attack rather than one-to-one | One-Pixel Signature detector · all-to-one attack | detection rate | No clear effect |
| default image constant 0 versus average testing image (ablation) | One-Pixel Signature · MNIST (LeNet-5) | detection rate | No clear effect |
Closest papers by meaning
- An Invisible Black-Box Backdoor Attack through Frequency Domain (ECCV 2022 · similarity 0.85)
- Detecting Backdoors in Pre-Trained Encoders (CVPR 2023 · similarity 0.85)
- A Dual Stealthy Backdoor: From Both Spatial and Frequency Perspectives (AAAI 2024 · similarity 0.83)
- Attention-Imperceptible Backdoor Attacks on Vision Transformers (AAAI 2025 · similarity 0.83)
- BackdoorDM: A Comprehensive Benchmark for Backdoor Learning on Diffusion Model (NeurIPS 2025 · similarity 0.83)
- Stealthy Backdoor Attack in Self-Supervised Learning Vision Encoders for Large Vision Language Models (CVPR 2025 · similarity 0.83)
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.
