Privacy Preserving Analytics in IoT Systems using Federated Learning with Homomorphic Encryption and Differential Privacy

In recent years, privacy preserving analytics has improved a lot which leads to development of smart platforms that include advanced mobile devices which are used by many people for businesses that are highly secured. These devices run apps that helps to improve online services which making things faster but some devices failed in securing data and maintaining privacy. To overcome those challenges, this research proposes a model namely Federated Learning with Homomorphic Encryption and Differential Privacy (FL-HE-DP). Initially, the data is collected from Telemetry and Network-Internet of Things (ToN-IoT) which contains nine different kinds of attacks, legitimate and offensive events in network systems. Then, the collected data is preprocessed using K-Nearest Neighbor (KNN) imputation, min-max scaling and one hot encoding which assists in handling noise, normalization and for converting categorical values into numerical values. Further, the preprocessed data is embedded into federated learning which utilizes HE for edge aggregations and encryption. Finally, the encrypted data is fed to DP, which helps to prevent data and adversarial attacks. The proposed FL-HE-DL acquired greater results by attaining Root Mean Squared Error (RMSE) of 8.965, Mean Absolute Error (MAE) of 7.654 and Mean Squared Error (MSE) of 5.764 respectively when compared to existing Soft Actor Critic (SAC) model.