New PromptMIA attack targets privacy in federated prompt-tuning

Researchers have developed a new privacy attack called PromptMIA, specifically designed for federated prompt-tuning scenarios. This attack exploits adversarially crafted prompts introduced by a malicious server during collaborative training to infer whether a target data point belongs to a client's private dataset. PromptMIA significantly outperforms existing federated membership inference attacks and demonstrates that current defenses are often ineffective, highlighting the need for specialized privacy mechanisms in federated prompt-tuning. AI

IMPACT Highlights a new privacy vulnerability in federated learning, necessitating the development of specialized defenses for prompt-tuning.

RANK_REASON The cluster contains a research paper detailing a novel privacy attack method. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• arXiv
• federated learning
• federated prompt-tuning
• PromptMIA
• Quan Nguyen

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →