Research map / Adversarial robustness and security / Backdoor and poisoning
Better Trigger Inversion Optimization in Backdoor Scanning
Guanhong Tao, Guangyu Shen, Yingqi Liu, Shengwei An, Qiuling Xu, Shiqing Ma, Pan Li, 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 |
|---|---|---|---|
| Replace NC's mask-and-perturbation optimization with direct per-pixel perturbation optimization using two tanh functions | ImageNet model trigger inversion · ImageNet validation, loggerhead turtle to kangaroo | ASR: 28.00% -> 70.00% | Improved |
| Replace NC's mask-and-perturbation optimization with direct per-pixel perturbation optimization | ImageNet model trigger inversion · ImageNet validation, loggerhead turtle to kangaroo | perturbed pixels: 26583 -> 822 | Improved |
| Use direct per-pixel perturbation optimization instead of NC | ImageNet model trigger inversion · ImageNet validation, loggerhead turtle to kangaroo | time (minutes): 9.02 -> 4.27 | Improved |
| Replace the full tanh pixel-change formulation in Equation 4 with the clipped linear formulation in Equation 6 (ablation) | Ours on clean models generating natural triggers · Natural trigger generation | trigger quality | No clear effect |
| Use the full tanh pixel-change formulation rather than the clipped linear formulation (ablation) | Ours on trojaned models during backdoor scanning · Backdoor scanning | trigger generation quality | Improved |
| Retain the tanh functions in the pixel-count penalty Lpixel (ablation) | Ours · Trigger generation | trigger quality | Improved |
| Use Ours instead of NC for trigger inversion | ImageNet model trigger inversion · ImageNet validation, loggerhead turtle to snowbird | ASR: 60.00% -> 72.00% | Improved |
| Use Ours instead of NC for trigger inversion | CIFAR-10 clean ResNet20 · CIFAR-10, all class pairs | ASR after matching trigger size: 39.83% -> 78.22% | Improved |
| Use Ours instead of NC to harden the model | ResNet32 · SVHN model hardening | average trigger-size increase: 152.02% | Improved |
| Replace the NC-based optimization component in K-arm with Ours | K-arm polygon-backdoor scanner · TrojAI competition models | detection accuracy: 2% | Improved |
| Use Ours instead of NC for inversion of WaNet backdoors | Backdoor scanning · WaNet | anomaly index | Improved |
Closest papers by meaning
- Towards Reliable and Efficient Backdoor Trigger Inversion via Decoupling Benign Features (ICLR 2024 · similarity 0.89)
- UNICORN: A Unified Backdoor Trigger Inversion Framework (ICLR 2023 · similarity 0.88)
- REFINE: Inversion-Free Backdoor Defense via Model Reprogramming (ICLR 2025 · similarity 0.88)
- BAN: Detecting Backdoors Activated by Adversarial Neuron Noise (NeurIPS 2024 · similarity 0.86)
- Backdoor Mitigation via Invertible Pruning Masks (NeurIPS 2025 · similarity 0.85)
- Trap-MID: Trapdoor-based Defense against Model Inversion Attacks (NeurIPS 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.
