Scholarlink

This paper explores the application of federated learning (FL) in human activity recognition (HAR), addressing the challenges posed by traditional centralized machine learning approaches that aggregate sensitive data on a central server. By leveraging FL, this study enables model training directly on user devices, thereby enhancing privacy and reducing communication costs. We specifically focus on the performance comparison between standard Long Short-Term Memory (LSTM) networks, a simplified version known as minimal LSTM (minLSTM), and Graph Neural Networks (GNNs) within a federated framework. The proposed methodology utilizes the FedAvg algorithm for aggregating locally trained models while ensuring data confidentiality. Our experiments, conducted using a comprehensive dataset of inertial sensor data from multiple participants, demonstrate that minLSTMs can achieve comparable performance to traditional LSTMs while significantly reducing computational overhead. Additionally, we show that GNNs effectively model structured sensor data, providing another viable approach for HAR. This research highlights the potential of FL to facilitate effective HAR in privacy-sensitive environments, paving the way for future advancements in decentralized machine learning applications.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Human Activity Recognition Using MinLSTM in a Federated Learning Approach

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Fatiha Barrade,
  • Moussa Jamor,
  • Mohamed Lazaar

摘要

This paper explores the application of federated learning (FL) in human activity recognition (HAR), addressing the challenges posed by traditional centralized machine learning approaches that aggregate sensitive data on a central server. By leveraging FL, this study enables model training directly on user devices, thereby enhancing privacy and reducing communication costs. We specifically focus on the performance comparison between standard Long Short-Term Memory (LSTM) networks, a simplified version known as minimal LSTM (minLSTM), and Graph Neural Networks (GNNs) within a federated framework. The proposed methodology utilizes the FedAvg algorithm for aggregating locally trained models while ensuring data confidentiality. Our experiments, conducted using a comprehensive dataset of inertial sensor data from multiple participants, demonstrate that minLSTMs can achieve comparable performance to traditional LSTMs while significantly reducing computational overhead. Additionally, we show that GNNs effectively model structured sensor data, providing another viable approach for HAR. This research highlights the potential of FL to facilitate effective HAR in privacy-sensitive environments, paving the way for future advancements in decentralized machine learning applications.