Researchers unveil BadVSFM, a new attack targeting video segmentation models

Researchers have developed a new method called BadVSFM to exploit vulnerabilities in prompt-driven video segmentation foundation models, such as SAM2. Traditional backdoor attacks were found to be ineffective against these models, achieving success rates below 5%. BadVSFM employs a two-stage approach to successfully implant triggers, leading to controllable backdoor effects with minimal degradation of clean performance. Existing defenses have proven largely ineffective against this new attack. AI

Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →

IMPACT New backdoor attack method highlights practical security risks in prompt-driven video segmentation models, potentially impacting their deployment in sensitive applications.

RANK_REASON Academic paper detailing a new attack method against existing AI models.

Read on arXiv cs.CV →

• BadVSFM
• SAM2
• Zongmin Zhang
• arXiv

• paper
• safety