Research map / Adversarial robustness and security / Attacks on text and graphs
FlowHijack: A Dynamics-Aware Backdoor Attack on Flow-Matching Vision-Language-Action Models
Xinyuan An, Tao Luo, Gengyun Peng, Yaobing Wang, Kui Ren, Dongxia Wang
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 |
|---|---|---|---|
| use FlowHijack dynamics-space attack instead of BadVLA with Object State trigger | pi0 · LIBERO-Goal | ASR: 11.2 -> 100 | Improved |
| use FlowHijack Ours(PL) instead of BadVLA with Scene Semantic trigger | pi0 · LIBERO-Spatial | ASR: 15.3 -> 100 | Improved |
| use Ours(PL) instead of BadVLA with White Pixel trigger | pi0 · LIBERO-10 | SR (no trigger): 72.9 -> 82.0 | Improved |
| add the flow matching loss L_FM (ablation) | FlowHijack · LIBERO suites | SR (no trigger): 0.0 -> 82.8 | Improved |
| add the backdoor hijacking loss L_BD (ablation) | FlowHijack · LIBERO suites | ASR: 0.0 -> 64.4 | Improved |
| add the dynamics mimicry regularizer L_mimic (ablation) | FlowHijack · LIBERO-10 | ASR: 66.7 -> 64.4 | Mixed |
| add the dynamics mimicry regularizer L_mimic (ablation) | FlowHijack · LIBERO-Object | SR (no trigger): 95.3 -> 98.8 | Improved |
| enlarge injection window tau0 from 0.1 to 0.4 (ablation) | Ours(IP) · LIBERO-Goal | ASR: 23.9 -> 100 | Improved |
| enlarge injection window tau0 from 0.4 to 0.8 (ablation) | Ours(IP) · LIBERO-Goal | SR (no trigger): 97.8 -> 96.6 | Hurt |
| tighten target position filter defense to 0.1 m | Ours(PL) · LIBERO-Goal | ASR: 100 -> 17.8 | Hurt |
| tighten target position filter defense to 0.1 m | Ours(IP) · LIBERO-Goal | ASR: 100 -> 82.2 | No clear effect |
| apply downstream clean LoRA fine-tuning for 10k steps | FlowHijack · LIBERO-Goal | ASR: 100 -> 67.7 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- BadVLA — BadVLA: Towards Backdoor Attacks on Vision-Language-Action Models via Objective-Decoupled Optimization (NeurIPS 2025) · better on 8 of 9 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- BadVLA: Towards Backdoor Attacks on Vision-Language-Action Models via Objective-Decoupled Optimization (NeurIPS 2025 · similarity 0.91)
- TRAP: Hijacking VLA CoT-Reasoning via Adversarial Patches (ICML 2026 · similarity 0.87)
- Image Hijacks: Adversarial Images can Control Generative Models at Runtime (ICML 2024 · similarity 0.87)
- Revisiting Backdoor Attacks against Large Vision-Language Models from Domain Shift (CVPR 2025 · similarity 0.86)
- Everyday Object Meets Vision-and-Language Navigation Agent via Backdoor (NeurIPS 2024 · similarity 0.86)
- FASTER: Rethinking Real-Time Flow VLAs (NeurIPS 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.
