Towards Probabilistic Verification of Machine Unlearning

• DOI: 10.48550/arxiv.2003.04247
• Publication Date: 2020-01-01
• Authors: David Sommer; Liwei Song; Sameer Wagh; Prateek Mittal
• Keywords: FOS: Computer and information sciences; Computer Science - Machine Learning; Computer Science - Cryptography and Security; Statistics - Machine Learning; 0202 electrical engineering, electronic engineering, information engineering; Machine Learning (stat.ML); 02 engineering and technology; 16. Peace & justice; Cryptography and Security (cs.CR); Machine Learning (cs.LG)

The right to be forgotten, also known as the erasure, is of individuals have their data erased from an entity storing it. status this long held notion was legally solidified recently by General Data Protection Regulation (GDPR) in European Union. Consequently, there a need for mechanisms whereby users can verify if service providers comply with deletion requests. In work, we take first step proposing formal framework study design such verification requests -- machine unlearning context systems that provide learning (MLaaS). Our allows rigorous quantification any mechanism based on standard hypothesis testing. Furthermore, propose novel backdoor-based and demonstrate its effectiveness certifying high confidence, thus providing basis quantitatively inferring unlearning. We evaluate our approach over range network architectures multi-layer perceptrons (MLP), convolutional neural networks (CNN), residual (ResNet), short-term memory (LSTM), well 5 different datasets. has minimal effect ML service's accuracy but provides confidence proposed works even only handful employ system ascertain compliance particular, just 5% participating, modifying half backdoor, merely 30 test queries, both false positive negative ratios below $10^{-3}$. show testing it against adaptive adversary uses state-of-the-art backdoor defense method.