Efficient Dropout-Resilient Aggregation for Privacy-Preserving Machine Learning

With the increasing adoption of data-hungry machine learning algorithms, personal data privacy has emerged as one key concerns that could hinder success digital transformation. As such, Privacy-Preserving Machine Learning (PPML) received much attention from both academia and industry. However, organizations are faced with dilemma that, on hand, they encouraged to share enhance ML performance, but other potentially be breaching relevant regulations. Practical PPML typically allows multiple participants individually train their models, which then aggregated construct a global model in privacy-preserving manner, e.g., based multi-party computation or homomorphic encryption. Nevertheless, most important applications large-scale PPML, by aggregating clients' gradients update for federated learning, such consumer behavior modeling mobile application services, some inevitably resource-constrained devices, may drop out system due mobility nature. Therefore, resilience aggregation become an problem tackled. In this paper, we propose scalable scheme can tolerate dropout at any time, is secure against semi-honest active malicious adversaries setting proper parameters. By replacing communication-intensive building blocks seed pseudo-random generator, relying additive property Shamir secret sharing scheme, our outperforms state-of-the-art schemes up 6.37$\times$ runtime provides stronger dropout-resilience. The simplicity makes it attractive implementation further improvements.