Research map / Adversarial robustness and security / Backdoor and poisoning
Attention Hijacking: Backdooring Text Dataset Distillation via Semantic Anchors
Hang Ren, Xin Wang, Tong Yue, Wen Chen, Junqing Le
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 |
|---|---|---|---|
| enforce attention hijacking (frozen one-hot attention label on trigger) instead of learnable attention labels | DI-Attn · AG News, trigger "the" | ASR (%): 43.05 -> 62.53 | Improved |
| use attention hijacking with semantic anchor trigger | DI-Attn · SST-2, trigger "film" | CTA (%): 85.78 -> 89.29 | Improved |
| use AH instead of DI-Attn with trigger "cf" | DI-Attn · SST-2, trigger "cf" | CTA (%): 88.51 -> 83.60 | Hurt |
| use AH with domain feature trigger "said" | DI-Attn · AG News, trigger "said" | CTA (%): 91.84 -> 91.08 | No clear effect |
| add attention-label guidance (ATTN-DD) to standard distillation (ablation) | clean distillation · SST-2 clean reference | CTA (%): 51.51 -> 85.02 | Improved |
| switch from standard distillation to attention-guided distillation for dynamic injection (ablation) | DI · SST-2, trigger "cf" | CTA (%): 63.49 -> 88.51 | Improved |
| static poisoning before distillation (SI) instead of dynamic injection | DI-Attn · AG News, trigger "cf" | ASR (%): 99.97 -> 2.39 (SI) | Hurt |
| increase attention loss weight gamma from 1.0 to 2.0 (ablation) | AH · SST-2, trigger "film" | CTA (%): CTA drops to 85.14% while ASR 99.91% | Hurt |
| increase poisoning ratio from 0.1% to 0.5% (ablation) | AH · SST-2, trigger "film" | ASR (%): saturation at 100% ASR with 0.5% | Improved |
| increase samples per class from 1 to 2 for the baseline (ablation) | DI-Attn · SST-2, trigger "film" | ASR (%) | Hurt |
| transfer distilled data to smaller victim models (BERT-L6, BERT-Tiny) (ablation) | AH · SST-2, trigger "film" | CTA and ASR | Improved |
| apply post-hoc clean fine-tuning (N=50) as defense | AH backdoored model · SST-2, trigger "film" | ASR (%): 99.39 -> 97.29 (CTA 89.29 -> 85.30) | No clear effect |
Closest papers by meaning
- Backdoor Attacks in Token Selection of Attention Mechanism (ICML 2025 · similarity 0.89)
- Backdoor Cleaning without External Guidance in MLLM Fine-tuning (NeurIPS 2025 · similarity 0.88)
- BackdoorLLM: A Comprehensive Benchmark for Backdoor Attacks and Defenses on Large Language Models (NeurIPS 2025 · similarity 0.87)
- CROW: Eliminating Backdoors from Large Language Models via Internal Consistency Regularization (ICML 2025 · similarity 0.87)
- Defense against Backdoor Attack on Pre-trained Language Models via Head Pruning and Attention Normalization (ICML 2024 · similarity 0.87)
- TCAP: Tri-Component Attention Profiling for Unsupervised Backdoor Detection in MLLM Fine-Tuning (ICML 2026 · similarity 0.86)
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.
