A new research paper titled "Merge Now, Regret Later: The Hidden Cost of Model Merging Is Adversarial Transferability" challenges the notion that model merging (MM) inherently provides adversarial robustness. The study, which involved extensive evaluations across eight MM methods, seven datasets, and six attack methods, found that MM cannot reliably defend against transfer attacks, with over 80% transfer rates observed. Key insights suggest that stronger MM methods and mitigating representation bias can increase vulnerability to transfer attacks, although weight averaging appears to be an exception. The findings offer practical guidance for designing secure machine learning systems that utilize model merging. AI
IMPACT Suggests that current model merging techniques may not provide the expected security benefits and could increase vulnerability to adversarial attacks.
RANK_REASON Research paper published on arXiv detailing findings about model merging and adversarial transferability. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
