Research map / Adversarial robustness and security / Backdoor and poisoning
Rethinking the Reverse-engineering of Trojan Triggers
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 feature-space hyperplane constraint instead of input-space-only constraints | FeatureRE · CIFAR-10 IA | detection accuracy: 55% -> 95% | Helped |
| same, on feature-space Filter attack | FeatureRE · CIFAR-10 Filter | detection accuracy: 83% -> 93% | Helped |
| same, on input-space BadNets | FeatureRE · CIFAR-10 BadNets | detection accuracy: 100% -> 98% | No clear effect |
| feature-space reverse-engineering vs trained classifiers | FeatureRE · VGG WaNet | detection accuracy: 0.53 -> 93% | Helped |
| break reverse-engineered hyperplane by neuron flipping | FeatureRE · CIFAR-10 WaNet mitigation | ASR: 98.17% -> 0.04% | Helped |
| split model at last conv layer instead of 9th layer (ablation) | FeatureRE · CIFAR-10 ResNet18 | detection accuracy: 86.50% -> 94.75% | Helped |
| vary tau1, tau2, tau3 constraint values (ablation) | FeatureRE · CIFAR-10 IA | detection accuracy: 0.90-0.95 | No clear effect |
| use fewer than 10 clean samples per class (ablation) | FeatureRE · CIFAR-10 IA | detection accuracy | Hurt |
| increase hyperplane size tau3 from 1% to 5% (ablation) | FeatureRE · CIFAR-10 WaNet mitigation | ASR: 57.75% -> 0.04% | Helped |
| vary clean samples 5 to 100 per class for mitigation (ablation) | FeatureRE · CIFAR-10 WaNet mitigation | ASR: 0.08% -> 0.06% | No clear effect |
| adaptive attacker breaks orthogonality of Trojan features (ablation) | FeatureRE · CIFAR-10 BadNets | detection accuracy: 98% -> 65% | Hurt |
Closest papers by meaning
- TrojanDec: Data-free Detection of Trojan Inputs in Self-supervised Learning (AAAI 2025 · similarity 0.86)
- Scanning Trojaned Models Using Out-of-Distribution Samples (NeurIPS 2024 · similarity 0.86)
- BaDExpert: Extracting Backdoor Functionality for Accurate Backdoor Input Detection (ICLR 2024 · similarity 0.85)
- Deep-TROJ: An Inference Stage Trojan Insertion Algorithm through Efficient Weight Replacement Attack (CVPR 2024 · similarity 0.84)
- REFINE: Inversion-Free Backdoor Defense via Model Reprogramming (ICLR 2025 · similarity 0.84)
- Rethinking Backdoor Attacks on Dataset Distillation: A Kernel Method Perspective (ICLR 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.