New LoMime method enables efficient privacy attacks on ML models

Researchers have developed a novel method called LoMime for performing membership inference attacks (MIAs) on machine learning models, even when only label-only access is available. This approach leverages model extraction, where a surrogate model is trained to mimic the target model's behavior. By using active sampling and synthetic data, LoMime significantly reduces the query costs typically associated with label-only MIAs, achieving comparable accuracy to state-of-the-art methods with a fraction of the queries. The framework has demonstrated effectiveness on tabular datasets and shows promise for extension to deep neural networks used in image recognition. AI

IMPACT Introduces a more efficient method for privacy attacks on ML models, potentially influencing future defensive strategies.

RANK_REASON Academic paper detailing a new methodology for membership inference attacks. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• Abdullah Çağlar Öksüz
• LoMime

• paper
• safety

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