Research map / Adversarial robustness and security / Backdoor and poisoning
Unmasking Backdoors: An Explainable Defense via Gradient-Attention Anomaly Scoring for Pre-trained Language Models
Anindya Sundar Das, Kangjie Chen, Monowar Bhuyan
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 combined attention and gradient anomaly scoring instead of attention-only or gradient-only (ablation) | X-GRAAD · AG's News (BadNets, RIPPLES; BERT, DistilBERT, ALBERT) | ASR | Improved |
| add combined attention and gradient anomaly scoring instead of single-signal scores (ablation) | X-GRAAD · AG's News (BadNets, RIPPLES; BERT, DistilBERT, ALBERT) | CACC | Improved |
| lower anomaly threshold percentile from 95% to 45% (ablation) | X-GRAAD on ALBERT · SST-2 (LWS attack) | ASR | Improved |
| lower anomaly threshold percentile from 95% to 45% (ablation) | X-GRAAD on ALBERT · SST-2 (LWS attack) | CACC | Improved |
| lower anomaly threshold percentile (ablation) | X-GRAAD on BERT and DistilBERT · SST-2 and IMDb | CACC | No clear effect |
| apply X-GRAAD defense to clean-label backdoored model | BERT · SST-2 (clean-label cf trigger) | ASR: 1.0 -> 0.012 | Improved |
| apply X-GRAAD defense to clean-label backdoored model | BERT · SST-2 (clean-label cf trigger) | accuracy on poisoned samples: 0.476 -> 0.910 | Improved |
| apply X-GRAAD to multi-token trigger 'the silver quill' | BERT · SST-2 (BadNets) | ASR: 1.00 -> 0.012 | Improved |
| apply X-GRAAD to multi-token trigger 'a crimson echo' | BERT · SST-2 (BadNets) | ASR: 1.00 -> 0.058 | Improved |
| apply X-GRAAD to multi-token trigger 'james bond' | BERT · SST-2 (BadNets) | ASR: 0.83 -> 0.022 | Improved |
| apply X-GRAAD to multi-token trigger 'the silver quill' | BERT · SST-2 (BadNets) | CACC: 0.918 -> 0.909 | Hurt |
| use X-GRAAD instead of PURE under BadPre domain shift | BERT · SST-2 (BadPre) | ASR: 0.760 -> 0.003 | Improved |
Closest papers by meaning
- Backdoor Token Unlearning: Exposing and Defending Backdoors in Pretrained Language Models (AAAI 2025 · similarity 0.92)
- Defense against Backdoor Attack on Pre-trained Language Models via Head Pruning and Attention Normalization (ICML 2024 · similarity 0.91)
- Probe before You Talk: Towards Black-box Defense against Backdoor Unalignment for Large Language Models (ICLR 2025 · similarity 0.89)
- BackdoorLLM: A Comprehensive Benchmark for Backdoor Attacks and Defenses on Large Language Models (NeurIPS 2025 · similarity 0.89)
- Backdoor Attacks in Token Selection of Attention Mechanism (ICML 2025 · similarity 0.89)
- Simulate and Eliminate: Revoke Backdoors for Generative Large Language Models (AAAI 2025 · similarity 0.88)
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.
