New POPS method recovers unlearned private data from MLLMs

Researchers have developed a new adversarial strategy called Prompt-Optimized Parameter Shaking (POPS) to recover unlearned multi-modality knowledge from Multimodal Large Language Models (MLLMs). This method aims to exploit vulnerabilities in existing Multi-modality Machine Unlearning (MMU) techniques, which are designed to remove private information. POPS works by optimizing prompts to elicit potential private examples from MLLMs and then using these synthesized outputs to fine-tune the models, thereby revealing sensitive information. Experiments indicate that POPS can significantly recover erased sensitive data, highlighting fundamental weaknesses in current MMU algorithms. AI

IMPACT Highlights potential vulnerabilities in current machine unlearning techniques, suggesting a need for more robust privacy protections in MLLMs.

RANK_REASON This is a research paper detailing a new method for recovering unlearned knowledge from MLLMs. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• arXiv
• Hugging Face
• MLLMs
• Multi-modality Machine Unlearning
• Prompt-Optimized Parameter Shaking

• paper
• safety

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