PRIVACY-PRESERVING IDENTITY PATTERN DISCOVERY USING FEDERATED LEARNING

Abstract

The increasing use of digital platforms, cloud services, and interconnected information systems has created a growing need for intelligent identity pattern discovery while maintaining the privacy and confidentiality of sensitive data. Conventional centralized data mining approaches require information from multiple sources to be collected in a common repository, increasing the risks of unauthorized access, data exposure, and privacy violations. Federated learning offers an alternative by enabling multiple organizations or distributed clients to collaboratively train machine learning models without directly exchanging their raw data. This study proposes a privacy-preserving framework for identity pattern discovery using federated learning. The proposed framework analyses identity-related behavioural characteristics, such as authentication events, login frequency, access patterns, and deviations from established user behaviour. Local models are trained at participating clients, while a central aggregation server combines model updates to develop a collaborative global model. Additional privacy mechanisms, including secure aggregation and differential privacy, are considered to reduce the risk of information leakage from model updates. The framework is designed to evaluate identity pattern discovery performance, privacy protection, communication overhead, and model robustness under distributed and heterogeneous data conditions. Performance can be assessed using precision, recall, F1-score, and detection accuracy, together with appropriate privacy and computational measures. The proposed approach aims to support trustworthy identity analytics in distributed environments, although its practical effectiveness must be established through empirical evaluation.