Dp-mtfl: Differentially Private Multi-Tier Federated Learning for IoT applications

Differentially Private Federated Learning (DP-FL) is a privacy-preserving machine learning paradigm. Building on a standard DP-FL approach, we introduce and implement a novel Differentially Private Multi-Tier Federated Learning approach specifically tailored for IoT applications, specifically short-term load forecasting. Our method integrates a Sampled Gaussian Mechanism for differential privacy with a hierarchical federated learning approach, where local federations participate in learning while adhering to approximate differential privacy with respect to a global server. We specifically study the optimal number of local rounds on global model convergence. Our findings demonstrate that non-DP models with fewer local rounds exhibit slightly superior performance compared to DP-enabled models. However, integrating DP by introducing additional noise during training with larger local rounds enhances the generalization of global models, suggesting that the sampled Gaussian mechanism functions as a form of regularization. In the evaluation of our method, we utilise an energy consumption dataset from the UK Power Networks Low Carbon London project. Our results show that our approach achieves privacy preserving objectives while obtaining the optimal number of local rounds that minimise the prediction error.