Research map / Adversarial robustness and security / Backdoor and poisoning
Practical Detection of Trojan Neural Networks: Data-Limited and Data-Free Cases
Ren Wang, Gaoyuan Zhang, Sijia Liu, Pin-Yu Chen, Jinjun Xiong, Meng Wang
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 |
|---|---|---|---|
| increase validation data per class from 1 to 5 (ablation) | DL-TND · CIFAR-10 | AUC: 0.96, 0.98 and 1 | Improved |
| increase data poisoning ratio from 0.5% to 10% (higher attack success) (ablation) | DL-TND · poisoning ratio sweep | AUC: 0.82 -> 0.99 | Improved |
| increase data poisoning ratio from 0.5% to 10% (higher attack success) (ablation) | DF-TND · poisoning ratio sweep | AUC: 0.7 -> 0.99 | Improved |
| use DL-TND instead of Neural Cleanse | detection of clean networks · Total of 85 models (T1 = 0.7) | correctly detected models: 54/85 -> 84/85 | Improved |
| use DL-TND instead of Neural Cleanse | detection of Trojan networks · Total of 85 models (T1 = 0.7) | correctly detected models: 48/85 -> 82/85 | Improved |
| use MAD detection rule instead of fixed threshold T1 = 0.7 (ablation) | DL-TND · Total of 85 models, clean networks | correctly detected models: 84/85 -> 67/85 | Hurt |
| use MAD detection rule instead of fixed threshold T1 = 0.7 (ablation) | DL-TND · Total of 85 models, Trojan networks | correctly detected models: 82/85 -> 74/85 | Hurt |
| use DL-TND instead of NC under MAD rule | detection of Trojan networks · Total of 85 models (MAD rule) | correctly detected models: 48/85 -> 74/85 | Improved |
| use random noise inputs instead of clean validation data as seeds (ablation) | DF-TND · CIFAR-10 model | AUC: 1 -> 0.99 | No clear effect |
| use Q = 0.75 instead of Q = 0.25 detection index (ablation) | DL-TND · Total, target label detection | AUC: 0.99 -> 0.98 | Hurt |
| replace DF-TND by DL-TND at low attack success (30% ASR, 0.5% poisoning) | detector choice · poisoning ratio 0.5% | AUC: 0.7 -> 0.82 | Improved |
| add refine step maximizing the Trojan-related coordinate (ablation) | DF-TND · recovered inputs | target label logits | Improved |
Closest papers by meaning
- Scanning Trojaned Models Using Out-of-Distribution Samples (NeurIPS 2024 · similarity 0.88)
- TrojanDec: Data-free Detection of Trojan Inputs in Self-supervised Learning (AAAI 2025 · similarity 0.88)
- Attack To Defend: Exploiting Adversarial Attacks for Detecting Poisoned Models (CVPR 2024 · similarity 0.87)
- Deep-TROJ: An Inference Stage Trojan Insertion Algorithm through Efficient Weight Replacement Attack (CVPR 2024 · similarity 0.86)
- Toward More Generalized Malicious URL Detection Models (AAAI 2024 · similarity 0.85)
- A Theoretical Analysis of Backdoor Poisoning Attacks in Convolutional Neural Networks (ICML 2024 · similarity 0.84)
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.
