Research map / Adversarial robustness and security / Backdoor and poisoning
Complex Backdoor Detection by Symmetric Feature Differencing
Yingqi Liu, Guangyu Shen, Guanhong Tao, Zhenting Wang, Shiqing Ma, Xiangyu Zhang
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 EX-RAY feature differencing after ABS trigger inversion | ABS · Composite attack across MNIST, FMNIST, SVHN, CIFAR10, and Youtube face | accuracy: 0.6 -> 0.88; 0.58 -> 0.83; 0.7 -> 0.88; 0.58 -> 0.85; 0.7 -> 0.85 | Improved |
| Add EX-RAY feature differencing after ABS trigger inversion | ABS · Reflection attack on CIFAR10 and ImageNet | accuracy: 0.65 -> 0.85; 0.63 -> 0.80; 0.55 -> 0.85; 0.68 -> 0.85; 0.59 -> 0.80; 0.62 -> 0.85 | Improved |
| Add EX-RAY feature differencing after ABS trigger inversion | ABS · TrojAI round 2 test set | ROC-AUC: 0.736 -> 0.892 | Improved |
| Add EX-RAY feature differencing after ABS trigger inversion | ABS · TrojAI round 3 test set | cross-entropy loss: 0.541 -> 0.323 | Improved |
| Add EX-RAY feature differencing after ABS trigger inversion | ABS · TrojAI round 4 test set | ROC-AUC: 0.549 -> 0.902 | Improved |
| Use symmetric rather than one-sided feature differencing (ablation) | EX-RAY · Backdoor detection | detection performance | Improved |
| Use target-class image features as the trigger, with trigger size 200 (ablation) | ABS+EX-RAY · Adaptive CIFAR10 attack | true-positive rate: 1 -> 0.5 | Hurt |
| Increase target-class-feature trigger size from 80 to 200 (ablation) | ABS+EX-RAY · Adaptive CIFAR10 attack | attack success rate on clean models: 0.39 -> 0.7 | Mixed |
Closest papers by meaning
- Towards Reliable and Efficient Backdoor Trigger Inversion via Decoupling Benign Features (ICLR 2024 · similarity 0.88)
- Lie Detector: Unified Backdoor Detection via Cross-Examination Framework (NeurIPS 2025 · similarity 0.87)
- Test-Time Backdoor Detection for Object Detection Models (CVPR 2025 · similarity 0.87)
- Activation Gradient based Poisoned Sample Detection Against Backdoor Attacks (ICLR 2025 · similarity 0.86)
- RPP: A Certified Poisoned-Sample Detection Framework for Backdoor Attacks under Dataset Imbalance (NeurIPS 2026 · similarity 0.86)
- UMA: Facilitating Backdoor Scanning via Unlearning-Based Model Ablation (AAAI 2024 · similarity 0.86)
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.
