Scholarlink

<p>Human Activity Recognition (HAR) aims to identify specific movements or actions based on sensor data. Existing techniques utilizing deep neural networks have several limitations, including insufficient accuracy in activity classification and a high computational cost. Moreover, centralized approaches that employ server-side analysis raise significant concerns regarding privacy. This research uses the transformer technique to address the first issue, leveraging the attention mechanism. This mechanism enables modeling dependencies regardless of the distance in the input or output sequences. The Transformer is not sensitive to the order of input tokens and lacks an inherent understanding of sequential positioning; thus, an additional encoding is added to the input tokens. However, this encoding can limit the model’s ability to capture fine-grained temporal relationships and may hinder optimization efficiency, leading to slower convergence. For this purpose, the proposed method adds positional encoding directly to the attention matrix at each head of the attention module. To address the second issue, Federated Learning (FL) is employed to maintain the privacy of user-sensitive data collected from wearable sensors. In our Federated Learning model, users only need to upload their local model weights to the server to generate training results, allowing sensitive data to remain on the user’s device. Experimental results demonstrate that MHAT-FL achieves state-of-the-art performance across PAMAP2, MotionSense, and Opportunity datasets, attaining F1-scores of 99.52%, 99.11%, and 99.79%, respectively. Moreover, MHAT-FL demonstrates strong privacy preservation by maintaining low Membership Inference Attack (MIA) accuracy, utilizing federated methods such as FedAvg and FedPer, with MIA accuracies of 16.67% on PAMAP2, 16.64% on MotionSense, and 16.66% on Opportunity. These results highlight MHAT-FL’s effectiveness in advancing Human Activity Recognition while ensuring strong privacy preservation, making it a reliable and secure solution for decentralized data scenarios.</p>

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 transformer-based positional encoding in a federated learning framework

  • 期刊论文

  • 发表日期：

    2025年9月18日

查看全文
  • Mohammad Ariaeimehr,
  • Reza Ravanmehr

摘要

Human Activity Recognition (HAR) aims to identify specific movements or actions based on sensor data. Existing techniques utilizing deep neural networks have several limitations, including insufficient accuracy in activity classification and a high computational cost. Moreover, centralized approaches that employ server-side analysis raise significant concerns regarding privacy. This research uses the transformer technique to address the first issue, leveraging the attention mechanism. This mechanism enables modeling dependencies regardless of the distance in the input or output sequences. The Transformer is not sensitive to the order of input tokens and lacks an inherent understanding of sequential positioning; thus, an additional encoding is added to the input tokens. However, this encoding can limit the model’s ability to capture fine-grained temporal relationships and may hinder optimization efficiency, leading to slower convergence. For this purpose, the proposed method adds positional encoding directly to the attention matrix at each head of the attention module. To address the second issue, Federated Learning (FL) is employed to maintain the privacy of user-sensitive data collected from wearable sensors. In our Federated Learning model, users only need to upload their local model weights to the server to generate training results, allowing sensitive data to remain on the user’s device. Experimental results demonstrate that MHAT-FL achieves state-of-the-art performance across PAMAP2, MotionSense, and Opportunity datasets, attaining F1-scores of 99.52%, 99.11%, and 99.79%, respectively. Moreover, MHAT-FL demonstrates strong privacy preservation by maintaining low Membership Inference Attack (MIA) accuracy, utilizing federated methods such as FedAvg and FedPer, with MIA accuracies of 16.67% on PAMAP2, 16.64% on MotionSense, and 16.66% on Opportunity. These results highlight MHAT-FL’s effectiveness in advancing Human Activity Recognition while ensuring strong privacy preservation, making it a reliable and secure solution for decentralized data scenarios.