Researchers have developed a new backdoor attack called ShadowFPT that targets Federated Prompt Tuning (FPT) in large vision-language models. This attack exploits the frozen backbone of the model, allowing a malicious client to inject backdoor behavior by optimizing a 'Shadow Trigger' against the frozen encoder. This trigger misdirects triggered inputs towards a target class in the representation space, while the client's prompt updates remain stealthy. ShadowFPT significantly increases attack success rates while maintaining clean accuracy, highlighting that defenses focused solely on prompt-update anomalies are insufficient. AI
IMPACT Highlights a new vulnerability in federated learning for LLMs, potentially requiring new defense mechanisms.
RANK_REASON Academic paper detailing a new attack method on AI models. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
