HAR Data Analysis: Unveiling the Potential of Federated CNN Models

The Internet of Things has seamlessly integrated into various facets of our daily lives, playing a pivotal role in sectors like security, transportation, smart homes, and healthcare. Across these domains, IoT employs numerous sensors that generate substantial data, traditionally stored in centralized infrastructures like Cloud Computing. However, this approach introduces challenges such as high latency, extended response times, and security concerns during data processing and analysis. In response to these challenges, federated learning, a revolutionary branch of artificial intelligence, has emerged. This paradigm shift allows machine learning techniques to benefit from both decentralized data and computing power, predominantly found in edge devices. This approach ensures a personalized user experience without compromising privacy. This paper introduces a federated Convolutional Neural Network (FedCNN) model and conducts a comparison between classical and federated learning for Human activities’ recognition. After a state of art, different aggregation approaches were evaluated including Avg-CNN, Sgd-CNN, Prox- CNN, Med-CNN, and KRUM-CNN. Avg-CNN model outperformed with an accuracy of 93.87% on the HAR dataset, while Prox- CNN Med-CNN, Sgd-CNN, and KRUM-CNN achieved respectively 93.77%, 93.38%, 93.38% and 85.40%.