MHAT-FL: An Open-Source TensorFlow Framework for Federated Human Activity Recognition with Attention-Matrix Positional Encoding

Abstract

Human Activity Recognition (HAR) systems deployed on wearable and edge devices continuously collect sensitive user data, raising critical privacy concerns and communication bottlenecks. While Federated Learning (FL) mitigates these issues by decentralizing model training and keeping raw data on-device, deploying highly accurate architectures such as Transformers within an FL ecosystem presents significant challenges. Traditional Transformer architectures, which add positional encoding directly to the input tokens, often struggle to capture the fine-grained temporal relationships inherent in high-frequency sensor data. In this paper, we introduce MHAT-FL, an open-source Python framework designed to streamline the deployment of Transformer-based models with attention-matrix positional encoding within a federated learning ecosystem. The software provides a modular architecture for simulating FL clients and servers, executing local training, and performing secure weight aggregation via Federated Personalization (FedPer). By offering reproducible pipelines and easy-to-use APIs without heavy third-party dependencies, MHAT-FL accelerates research in edge AI, TinyML, and privacy-preserving wearable computing.