Hybrid Privacy-Preserving Federated Learning Framework for Secure IoT Applications Using Differential Privacy and Homomorphic Encryption

Abstract

Internet of Things (IoT) applications have expanded quickly and are producing massive amounts of data from a diverse range of smart devices. Federated Learning (FL) allows collaborative model training without data transfer of the actual images but suffers from privacy concerns as sensitive information might be leaked through the model parameters. Most of the current privacy-preserving FL methods use DP or HE separately which leads to a compromise between privacy protection, model accuracy and computation efficiency. In response, this paper introduces Hybrid Privacy-Preserving Federated Learning (HPPFL), a comprehensive privacy-preserving learning system that combines Differential Privacy (DP) and Homomorphic Encryption (HE) for secure IoT applications. The proposed design allows for local training of the model on IoT devices, followed by the addition of noise (adaptive Differential Privacy) to the model updates, then encrypting the noisy model using Paillier Homomorphic Encryption before sending it to the federated server. The model updates are then securely aggregated with FedAvg without making sensitive information public and the aggregated model is sent back to the global model for next training round. TON-IoT was used to test the proposed framework, and it was compared to the conventional FL, DP-FL, and HE-FL approaches based on classification accuracy, precision, recall, F1-score, communication cost, encryption overhead, and training time. The experimental results demonstrate that the proposed HPPFL framework achieved classification accuracy of 97.46% compared to other FL frameworks, conventional FL, DP-FL, HE-FL with the highest level of privacy protection and moderate computational overhead. The results indicate that the proposed framework is capable of preserving privacy, ensuring effective communication and maintaining the prediction performance, which shows it is a potential solution to secure large-scale IoT applications.