Research map / Adversarial robustness and security / Backdoor and poisoning
Reflection Backdoor: A Natural Backdoor Attack on Deep Neural Networks
Yunfei Liu, Xingjun Ma, James Bailey, Feng Lu
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 |
|---|---|---|---|
| add ghost-effect (III) reflection model instead of same-DOF (I) (ablation) | Refool · GTSRB | attack success rate (%): 87.30% -> 90.33% | Improved |
| add out-of-focus (II) reflection model instead of same-DOF (I) (ablation) | Refool · GTSRB | attack success rate (%): 87.30% -> 90.46% | Improved |
| use random mix of three reflection types instead of fixed type (II) (ablation) | Refool · GTSRB | attack success rate (%): 90.46% -> 91.67% | Improved |
| use ghost-effect reflection instead of out-of-focus reflection (ablation) | Refool · GTSRB | SSIM image similarity: 0.896 -> 0.786 | Hurt |
| increase in-class injection rate up to 40% (ablation) | Refool · GTSRB | attack success rate | Improved |
| increase in-class injection rate beyond 40% (ablation) | Refool · GTSRB | attack success rate | No clear effect |
| use Refool instead of CL under fine-tuning defense | Refool · GTSRB | attack success rate after finetuning: CL 78.3% -> 20% after one epoch; Refool above 60% after 15 epochs | Improved |
| use Refool instead of CL under neural pruning | Refool · GTSRB | attack success rate under fine-pruning: CL 76% -> 8.3% at 60% neurons removed | Improved |
| white-box reflection removal applied to Refool | Refool · GTSRB | attack success rate: 91.67 -> 85.01 | Hurt |
| white-box trigger removal applied to CL | CL · GTSRB | attack success rate: 78.03 -> 18.18 | Hurt |
| white-box trigger removal applied to Refool on ImageNet with ResNet | Refool · ImageNet subset | attack success rate: 82.11 -> 36.93 | Hurt |
| apply image quilting defense | Refool · CTSRD | test accuracy (%): 86.30% -> 11.35% | Hurt |
Closest papers by meaning
- Architectural Backdoors in Neural Networks (CVPR 2023 · similarity 0.89)
- Training with More Confidence: Mitigating Injected and Natural Backdoors During Training (NeurIPS 2022 · similarity 0.89)
- BAN: Detecting Backdoors Activated by Adversarial Neuron Noise (NeurIPS 2024 · similarity 0.89)
- REFINE: Inversion-Free Backdoor Defense via Model Reprogramming (ICLR 2025 · similarity 0.88)
- BaDExpert: Extracting Backdoor Functionality for Accurate Backdoor Input Detection (ICLR 2024 · similarity 0.87)
- BackdoorDM: A Comprehensive Benchmark for Backdoor Learning on Diffusion Model (NeurIPS 2025 · similarity 0.87)
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.
