Research map / Adversarial robustness and security / Backdoor and poisoning
Detecting Backdoors During the Inference Stage Based on Corruption Robustness Consistency
Xiaogeng Liu, Minghui Li, Haoyu Wang, Shengshan Hu, Dengpan Ye, Hai Jin, Libing Wu, Chaowei Xiao
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 corruption-severity transition dispersion with hard-label outputs for test-time trigger detection | TeCo · CIFAR10, PreActResNet18 | AUROC: 0.584 -> 0.918 (STRIP); 0.888 -> 0.918 (FreqDetector) | Improved |
| Use corruption-severity transition dispersion with hard-label outputs for test-time trigger detection | TeCo · GTSRB, PreActResNet18 | AUROC: 0.690 -> 0.940 (STRIP); 0.784 -> 0.940 (FreqDetector) | Improved |
| Increase maximum corruption severity N (ablation) | TeCo · Multiple datasets and CNN/ViT backbones | AUROC | Mixed |
| Vary the predefined target label (ablation) | TeCo · GTSRB: Blended, SSBA, and Wanet attacks | AUROC: Blended AUROC ≥0.874, STD about 0.021; SSBA AUROC ≥0.908, STD ≈0.010; Wanet AUROC ≥0.919, STD ≈0.017 | No clear effect |
| Add an adaptive loss that aligns corruption robustness of clean and trigger inputs to backdoor training (ablation) | Backdoor models evaluated by TeCo · CIFAR10, PreActResNet18, BadNets | AUROC: 0.9112 -> 0.5763 (10^-3); 0.9112 -> 0.6571 (10^-4); 0.9112 -> 0.6745 (10^-5) | Hurt |
| Add an adaptive loss that aligns corruption robustness of clean and trigger inputs to backdoor training (ablation) | Backdoor models evaluated by TeCo · CIFAR10, PreActResNet18, SSBA | AUROC: 0.8683 -> 0.7125 (10^-3); 0.8683 -> 0.6477 (10^-4); 0.8683 -> 0.5909 (10^-5) | Hurt |
| Add an adaptive loss that aligns corruption robustness of clean and trigger inputs to backdoor training (ablation) | Backdoor models evaluated by TeCo · CIFAR10, PreActResNet18, LF | AUROC: 0.9390 -> 0.8592 (10^-3); 0.9390 -> 0.9219 (10^-4); 0.9390 -> 0.8667 (10^-5) | Mixed |
| Use one empirical threshold γ=1 across attacks (ablation) | TeCo · Multiple datasets and CNN/ViT backbones | accuracy: CNN AVG 0.7923; ViT AVG 0.7868 | Mixed |
Closest papers by meaning
- Lie Detector: Unified Backdoor Detection via Cross-Examination Framework (NeurIPS 2025 · similarity 0.89)
- Test-Time Backdoor Detection for Object Detection Models (CVPR 2025 · similarity 0.88)
- RPP: A Certified Poisoned-Sample Detection Framework for Backdoor Attacks under Dataset Imbalance (NeurIPS 2026 · similarity 0.88)
- IBD-PSC: Input-level Backdoor Detection via Parameter-oriented Scaling Consistency (ICML 2024 · similarity 0.87)
- Inspecting Prediction Confidence for Detecting Black-Box Backdoor Attacks (AAAI 2024 · similarity 0.87)
- LOTUS: Evasive and Resilient Backdoor Attacks through Sub-Partitioning (CVPR 2024 · 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.
