Privacy-Preserving Data Mining Using Federated Learning and Differential Privacy

Abstract

The goal of privacy-preserving data mining is to discover useful patterns in the distributed data, while minimizing the disclosure of individual sensitive information. Federated learning meets this challenge by training a shared model without collecting raw data at a central location, while differential privacy gives a formal guarantee of a bounded influence of any individual's participation in the model outputs published. This paper summarizes the state-of-the-art of the application of federated learning and differential privacy in privacy-preserving data mining, summarizes some fundamental algorithms, and summarizes recent research directions, and proposes a reproducible method to evaluate the privacy–utility–communication trade-off. The analysis shows that significant issues are encountered, such as statistical heterogeneity, privacy-budget accumulation, malicious participants, communication overhead, and limited transparency in privacy accounting. The paper shows that the rise of data mining applications in healthcare, finance, mobile computing, and public-sector analytics demands the use of practical systems that leverage decentralized training, client-level differential privacy, secure aggregation, strong privacy accounting, and extensive empirical testing to ensure responsible mining.