Differential Privacy in Secure Multiparty Computation and Deep Neural Networks
Abstract
Differential privacy (DP), a leading framework in data privacy research, ensures that individual data points remain confidential, even when aggregate information is shared. It provides a quantifiable level of privacy protection, balancing privacy risks with data utility. However, putting DP into practice still faces challenges, requiring further developments to effectively balance privacy, data utility, and scalability.
This doctoral thesis explores the implementation of DP in secure multiparty computation (MPC) and its application in machine learning (ML). The integration of DP with MPC seeks to enable collaborative computations on sensitive data in a distributed setting, while the combination of DP with ML aims to enhance privacy in a centralized model training period.
The thesis is composed of six research papers. Two papers introduce MPC protocols with different security guarantees for sampling a biased coin, an important component of implementing DP.
Two other papers describe an MPC framework for DP, focusing on handling floatingpoint approximation of real-valued queries while preserving DP guarantees. The framework consists of three protocols to compute the Laplace mechanism specifically for linear queries, which differ in communication cost, round complexity, and probability of failure.
The final two papers investigate the application of DP in ML, particularly in image classification and face recognition models. These studies examine the trade-offs between privacy, accuracy, and fairness, providing insights into the challenges of applying DP in practical ML contexts.
Has parts
Paper A: Zarei, Amir; Vinterbo, Staal Amund. Secure Multiparty Sampling of a Biased Coin for Differential Privacy. Lecture Notes in Computer Science , vol 14398. (LNCS) 2024 https://doi.org/10.1007/978-3-031-54204-6_19Paper B: Zarei, Amir; Vinterbo, Staal Amund. Secure Multiparty Computation of the Laplace Mechanism. ICISSP 2024 s. 582-593 Proceedings of the 10th International Conference on Information Systems Security and Privacy ICISSP - Volume 1 https://doi.org/10.5220/0012453700003648 CC BY-NC-ND
Paper C: Zarei, Amir; Vinterbo, Staal Amund. Statistically secure multiparty computation of a biased coin. Security and Trust Management: 20th International Workshop, STM 2024 - Lecture Notes in Computer Science LNCS,volume 15235 https://doi.org/10.1007/978-3-031-76371-7_5
Paper D: Amir; Vinterbo, Staal Amund. Hole-Free Differentially Private Multiparty Laplace Mechanism
Paper E: The Impact of Generalization Techniques on the Interplay Among Privacy, Utility, and Fairness in Image Classification https://doi.org/10.48550/arXiv.2412.11951 CC BY
Paper F: Zarei, Amir; Hassanpour, Ahmad; Raja, Kiran. On Privacy, Accuracy, and Fairness Trade-offs in Facial Recognition