Federated Learning with Smartphone Sensor Data

Federated learning (FL) is emerging as a promising approach for training machine learning models on distributed devices without violating the data privacy of these devices. In this paper, we examine federated learning for smartphone sensor data applications that involve both significant challenges related to privacy and data heterogeneity. Our primary interest lies in techniques such as differential privacy, secure aggregation, and homomorphic encryption that will ensure privacy over user-sensitive information during model training. We also pose and discuss heterogeneous data across devices, particularly non-independent and identically distributed (non-IID) data, by investigating methods such as normalization of data, personalized learning, and federated transfer learning. Using real-world smartphone sensor datasets, we demonstrate experimentally that federated learning is effective in training robust models while preserving privacy and accounting for device-specific data variations. Our findings highlight that federated learning can be regarded as a way to scale-up and privacy-preserve mobile-based machine learning, which may open new avenues for building real-time AI systems on t