Research map / Adversarial robustness and security / Attacks on text and graphs
VLA-Hijack: A Transferable Patch Attack against Vision-Language-Action Models via Visual Proprioception Hijacking
Jiyuan Fu, Kaixun Jiang, Jingkai Jia, Zhaoyu Chen, Xueyao Chen, Lingyi Hong, Shuyong Gao, Chenzhi Tan, Dingkang Yang, Wenqiang Zhang
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 |
|---|---|---|---|
| alternate semantic and visual injection (tau=5) instead of joint optimization (ablation) | VLA-Hijack · UniVLA variants avg | failure rate: 66.25 -> 74.25 | Improved |
| alternating period 3 vs 5 vs 10 (ablation) | VLA-Hijack · UniVLA variants avg | failure rate: 67.50 / 74.25 / 72.00 | Mixed |
| injection weight lambda 0.1 -> 0.2 (ablation) | VLA-Hijack · UniVLA variants avg | failure rate: 64.25 -> 74.25 | Improved |
| increase lambda 0.2 -> 0.5 (ablation) | VLA-Hijack · UniVLA variants avg | failure rate: 74.25 -> 66.00 | Hurt |
| add attention-guided suppression term (ablation) | VLA-Hijack · UniVLA variants avg | failure rate | Improved |
| add semantic anchoring and visual prototype injection (ablation) | VLA-Hijack · UniVLA variants avg | failure rate | Improved |
| VLA-Hijack vs best action-space baseline on cross-architecture transfer | VLA-Hijack · LIBERO 11 black-box victims | failure rate: 37.53 -> 61.64 | Improved |
| VLA-Hijack vs baselines under Real-to-Sim transfer | VLA-Hijack · LIBERO 12 victims (Real-to-Sim) | failure rate: 21.02 -> 63.68 | Improved |
| naive arm image patch vs optimization-based attacks (Real-to-Sim) | Arm image · LIBERO 12 victims (Real-to-Sim) | failure rate: 18.68 (UADA_1-3) -> 23.45 | Improved |
| action-space attack white-box (UADA_1-3) vs transfer | UADA_1-3 · LIBERO | failure rate: 100.00 (white-box) -> 37.53 (transfer avg) | Hurt |
| UniVLA-Object as surrogate instead of OpenVLA-Spatial | VLA-Hijack · LIBERO transfer avg | failure rate: 61.64 -> 64.35 | Improved |
| VLA-Hijack white-box convergence speed | VLA-Hijack · OpenVLA-Spatial | steps to 100% FR: 30 steps | Improved |
Closest papers by meaning
- BadVLA: Towards Backdoor Attacks on Vision-Language-Action Models via Objective-Decoupled Optimization (NeurIPS 2025 · similarity 0.90)
- TRAP: Hijacking VLA CoT-Reasoning via Adversarial Patches (ICML 2026 · similarity 0.90)
- VLA-Corrector: Lightweight Detect-and-Correct Inference for Adaptive Action Horizon (NeurIPS 2026 · similarity 0.89)
- ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.89)
- VLA-OS: Structuring and Dissecting Planning Representations and Paradigms in Vision-Language-Action Models (NeurIPS 2025 · similarity 0.88)
- VLA-Cache: Efficient Vision-Language-Action Manipulation via Adaptive Token Caching (NeurIPS 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.
