Research map / Adversarial robustness and security / Certified and empirical robustness
Towards Effective and Robust Neural Trojan Defenses via Input Filtering
Kien Do, Haripriya Harikumar, Hung Le, Dung Nguyen, Truyen Tran, Santu Rana, Dang Nguyen, Willy Susilo, Svetha Venkatesh
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 KL (lossy compression) term, IF -> VIF (ablation) | IF · CIFAR10 (image-BI+) | Trojan accuracy T (%): 37.48 -> 10.52 | Improved |
| add adversarial trigger generator, IF -> AIF (ablation) | IF · CIFAR10 (image-BI+) | Trojan accuracy T (%): 37.48 -> 1.96 | Improved |
| add KL term, IF -> VIF (ablation) | IF · GTSRB (image-BI+) | Trojan accuracy T (%): 52.27 -> 12.12 | Improved |
| add adversarial training, IF -> AIF (ablation) | IF · CelebA (image-BI+) | Trojan accuracy T (%): 59.39 -> 18.40 | Improved |
| add KL term, IF -> VIF (ablation) | IF · CIFAR10 (benign model) | decrease in clean accuracy (%): 3.34 -> 7.81 | Hurt |
| add adversarial training, IF -> AIF | VIF · CIFAR10 (benign model) | decrease in clean accuracy (%): 7.81 -> 4.67 | Improved |
| add filtering-then-contrasting to AIF | AIF · CIFAR10 (image-BI+) | FPR/FNR vs T: 1.96 -> 2.15 | No clear effect |
| add FtC to filtering defense for large-norm BadNet+ triggers (s=23) (ablation) | IF/VIF · CIFAR10 (BadNet+ s=23) | Trojan detection FNR (%): 64.41 -> 63.74 | Improved |
| use Filtering-then-Contrasting instead of filtering under large-norm triggers (ablation) | VIF · CIFAR10 (noise-BI+ alpha=0.7) | FNR (%): 0.22 -> 0.26 | Improved |
| reduce training data with fixed epochs (proportion 1.0 -> 0.05) (ablation) | IF · CIFAR10 (InpAwAtk) | decrease in clean accuracy (%): 4.47 -> 25.93 | Hurt |
| reduce training data with fixed total iterations (1.0 -> 0.05) (ablation) | VIF · CIFAR10 (InpAwAtk) | Trojan accuracy T (%): 3.07 -> 4.67 | No clear effect |
| remove explicit trigger normalization in AIF (AIF*) (ablation) | AIF · CIFAR10 (BadNet+ s=23) | FNR (%): 28.0 -> 30.19 | Hurt |
Closest papers by meaning
- Practical Detection of Trojan Neural Networks: Data-Limited and Data-Free Cases (ECCV 2020 · similarity 0.90)
- BppAttack: Stealthy and Efficient Trojan Attacks Against Deep Neural Networks via Image Quantization and Contrastive Adversarial Learning (CVPR 2022 · similarity 0.88)
- Scanning Trojaned Models Using Out-of-Distribution Samples (NeurIPS 2024 · similarity 0.85)
- TrojanDec: Data-free Detection of Trojan Inputs in Self-supervised Learning (AAAI 2025 · similarity 0.85)
- Attack To Defend: Exploiting Adversarial Attacks for Detecting Poisoned Models (CVPR 2024 · similarity 0.85)
- Deep-TROJ: An Inference Stage Trojan Insertion Algorithm through Efficient Weight Replacement Attack (CVPR 2024 · similarity 0.85)
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.
