Research map / Adversarial robustness and security / Backdoor and poisoning
Finding Naturally Occurring Physical Backdoors in Image Datasets
Emily Wenger, Roma Bhattacharjee, Arjun Nitin Bhagoji, Josephine Passananti, Emilio Andere, Haitao Zheng, Ben Zhao
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 the approximate MIS constraint when choosing the poisonable class subset (ablation) | centrality-based natural backdoor dataset · Open Images | clean accuracy: 58 ± 5% -> 72 ± 1% | Improved |
| add the approximate MIS constraint when choosing the poisonable class subset (ablation) | centrality-based natural backdoor dataset · Open Images | trigger accuracy: 63 ± 8% -> 68 ± 3% | Improved |
| natural backdoor dataset versus training without any backdoor | Open Images models · Open Images | clean accuracy: 79 ± 2% -> 72 ± 1% | Hurt |
| switch model architecture from VGG16 to ResNet (ablation) | natural backdoor jeans dataset · Open Images jeans trigger | trigger accuracy: 69 ± 5% -> 75 ± 4% | Improved |
| switch model architecture from ResNet to Inception (ablation) | natural backdoor jeans dataset · Open Images jeans trigger | trigger accuracy: 75 ± 4% -> 61 ± 1% | Hurt |
| vary model architecture (DenseNet, ResNet, VGG16, Inception) (ablation) | natural backdoor jeans dataset · Open Images jeans trigger | clean and trigger accuracy | No clear effect |
| increase poison injection rate (ablation) | natural backdoor jeans dataset · Open Images jeans trigger | clean and trigger accuracy | Improved |
| use betweenness centrality instead of other centrality measures (ablation) | trigger selection · Open Images and ImageNet | clean/trigger accuracy consistency | Improved |
| use closeness centrality instead of other centrality measures (ablation) | trigger selection · Open Images | number of viable triggers | Hurt |
| remove spectral-flagged data and retrain | natural backdoor models · natural backdoor datasets | attack accuracy: 4% | No clear effect |
| apply NeuralCleanse, Activation Clustering, Spectral, STRIP defenses | natural backdoor models · natural backdoor datasets | defense performance: 16%, 9.7 ± 10.8%, 65.0 ± 4.3%, 4.0 ± 4.0% | No clear effect |
| apply SentiNet defense (saliency maps) | natural backdoor models · natural backdoor datasets | defense performance: 56.9 ± 18.5% | Improved |
Closest papers by meaning
- Clean Data Can Still Carry Backdoors: Support-Persistent Backdoors for Model Reuse (NeurIPS 2026 · similarity 0.88)
- Universal Backdoor Attacks (ICLR 2024 · similarity 0.87)
- Rethinking Backdoor Attacks on Dataset Distillation: A Kernel Method Perspective (ICLR 2024 · similarity 0.87)
- Demystifying Poisoning Backdoor Attacks from a Statistical Perspective (ICLR 2024 · similarity 0.87)
- Protecting against simultaneous data poisoning attacks (ICLR 2025 · similarity 0.87)
- Invisible Backdoor Attack against Self-supervised Learning (CVPR 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.
