Homomorphic Encryption based Federated Learning for Financial Data Security

Federated Learning is a distributed machine learning technique that enables on-device training without exchanging the sensitive data over the centralized server. In this paper, Federated Learning is used to train financial data models on-devices with the help of IoT applications in financial or business systems. This creates more advanced and secured financial business models. We have applied the mechanism of homomorphic encryption and cryptographic primitives, including masking and local model protection to prevent any kind of inferring in private financial or business data where multiple attackers usually find a way of model inversion or reconstruction attack. We used datasets of various financial sectors as the primary measure, rather than taking the size of datasets that is generally used in deep learning, to get the correct measurement of the rate of contribution in every session of the local model’s training to the training of the global model. If the calculated online number of clients exceeds a predetermined threshold, then the federated learning process will be continued with the dropout-tolerant plan. The security study demonstrates that the suggested solution fulfils data privacy requirements. The costs of computation and communication are also examined theoretically. According to research observations, the proposed approach achieved promising outcomes while assuring privacy preservation when compared to existing schemes.