BEAT: Visual Backdoor Attack Training Data for Embodied AI Agents

BEAT is a training dataset for backdoor attacks on vision-language model-based embodied agents, accompanying a 2026 arXiv paper. The dataset is gated and intended for academic research on AI security and robustness. It was created by the UIUC Kang Lab and last updated on April 20, 2026.

Use Cases

• Training and evaluating backdoor attack methods based on the described contrastive trigger learning approach.
• Researching the security vulnerabilities of vision-language models in embodied agent scenarios.
• Developing defense mechanisms against visual backdoor attacks in multimodal AI systems.

Strengths

• Dataset is associated with a specific, recent 2026 research paper (arXiv:2510.27623).
• Access is controlled, suggesting a defined research purpose and potential for responsible use.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count and file formats are unknown, which may limit suitability assessment.

Provenance

Source

uiuc-kang-lab

Collection Method

Created to accompany the research paper 'BEAT: Visual Backdoor Attacks on VLM-based Embodied Agents via Contrastive Trigger Learning'.

Freshness

Last updated 2026-04-20 06:12:32; freshness should be verified.