Scholarlink

This study presents a privacy-preserving Sign Language Recognition (SLR) system using Dynamic Weighted Federated Learning (DWFL), addressing critical limitations of centralized approaches. Conventional SLR systems face substantial privacy risks due to centralized data processing, while our federated learning (FL) framework, enables collaborative training across devices without sharing raw data. The proposed solution combines Vision-based CNNs, and sensor-based methods for static/dynamic sign recognition, DWFL optimization that prioritizes high-quality local models during aggregation, Differential privacy (ε < 1.5) for GDPR-compliant data security. Experimental results on the ASLLVD dataset demonstrate 92.3% recognition accuracy (vs. 85.7% for FedAvg), <200 ms latency for real-time translation, Robust performance across non-IID data distributions. This work establishes a foundation for secure assistive technologies that bridge communication gaps for deaf/hard-of-hearing communities while addressing scalability challenges.

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

Privacy-Preserving Sign Language Recognition: A Federated Learning Approach for Decentralized and Secure Communication Systems

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Adarsh Kumar Dubey,
  • Aditya Raj,
  • Adrisha Gupta,
  • Anas Aijaz,
  • Meeta Chaudhry

摘要

This study presents a privacy-preserving Sign Language Recognition (SLR) system using Dynamic Weighted Federated Learning (DWFL), addressing critical limitations of centralized approaches. Conventional SLR systems face substantial privacy risks due to centralized data processing, while our federated learning (FL) framework, enables collaborative training across devices without sharing raw data. The proposed solution combines Vision-based CNNs, and sensor-based methods for static/dynamic sign recognition, DWFL optimization that prioritizes high-quality local models during aggregation, Differential privacy (ε < 1.5) for GDPR-compliant data security. Experimental results on the ASLLVD dataset demonstrate 92.3% recognition accuracy (vs. 85.7% for FedAvg), <200 ms latency for real-time translation, Robust performance across non-IID data distributions. This work establishes a foundation for secure assistive technologies that bridge communication gaps for deaf/hard-of-hearing communities while addressing scalability challenges.