Researchers have developed a new adversarial strategy called Prompt-Optimized Parameter Shaking (POPS) that can recover supposedly unlearned multi-modality knowledge from Multimodal Large Language Models (MLLMs). This method aims to address vulnerabilities in 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 recovering the erased sensitive information. Experiments show that POPS can significantly recover information even from models that have undergone unlearning, highlighting fundamental weaknesses in current MMU algorithms. AI
IMPACT Reveals fundamental vulnerabilities in current AI unlearning techniques, potentially impacting privacy and copyright protections for multimodal models.
RANK_REASON The cluster contains a research paper detailing a new method for recovering unlearned data from multimodal AI models. [lever_c_demoted from research: ic=1 ai=1.0]
Read on Hugging Face Daily Papers →
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
