Research map / Multi-agent systems and games / LLM agents
SlowBA: An efficiency backdoor attack towards VLM-based GUI agents
Junxian Li, Tu Lan, Haozhen Tan, Yan Meng, Haojin Zhu
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 Stage II trigger-aware reward optimization on top of Stage I format alignment (ablation) | SlowBA · Web | I-length (%): 6.54 -> 358.52 | Improved |
| add Stage I long-response format alignment before Stage II RL (ablation) | SlowBA · Web | I-latency (%): -59.74 -> 66.92 | Improved |
| add Stage I alignment before Stage II (ablation) | SlowBA · Web | I-length (%): 7.71 -> 358.52 | Improved |
| add Stage II to Stage I (trigger separation) (ablation) | SlowBA · Web | triggered length (tokens): 829.62 vs 778.66 -> 722.45 vs 157.56 | Improved |
| scale model from GUI-R1-3B to 7B | SlowBA · Web | I-latency (%): 66.92 -> 103.47 | Improved |
| scale model from GUI-R1-3B to 7B | SlowBA · Web | I-length (%): 358.52 -> 242.00 | Hurt |
| apply JPEG compression defense | SlowBA · Web | I-length (%): 358.52 -> 325.71 | Hurt |
| apply re-training on clean data defense | SlowBA · Web | I-length (%): 358.52 -> 275.36 | Hurt |
| apply Beatrix detection defense | SlowBA · Web | I-length (%): 358.52 -> 375.28 | No clear effect |
| apply Spectral Signature detection defense | SlowBA · Web | I-length (%): 358.52 -> 351.01 | No clear effect |
| backdoor only the MLP projector instead of the LLM (ablation) | SlowBA · Web | I-latency (%): 66.92 -> 59.17 | Hurt |
| backdoor only the visual encoder instead of the LLM (ablation) | SlowBA · Web | I-latency (%): 66.92 -> 70.75 | Improved |
Closest papers by meaning
- AdapAction: Adaptive Target Action Backdoor Attack against GUI Agents (CVPR 2026 · similarity 0.89)
- GUI-G1: Understanding R1-Zero-Like Training for Visual Grounding in GUI Agents (NeurIPS 2025 · similarity 0.87)
- BadVLA: Towards Backdoor Attacks on Vision-Language-Action Models via Objective-Decoupled Optimization (NeurIPS 2025 · similarity 0.87)
- GUI-Shift: Enhancing VLM-Based GUI Agents through Self-supervised Reinforcement Learning (ICLR 2026 · similarity 0.86)
- Faithful Mobile GUI Agents with Guided Advantage Estimator (ICML 2026 · similarity 0.86)
- SE-GUI: Enhancing Visual Grounding for GUI Agents via Self-Evolutionary Reinforcement Learning (NeurIPS 2025 · 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.
