Scholarlink

Electroencephalography (EEG)-based Brain – Computer Interface (BCI) systems enable direct communication between the brain and external devices, facilitating neurorehabilitation. Motor imagery (MI) classification and event-related potential (ERP) detection are two critical paradigms for developing efficient EEG-based BCIs. While deep learning enhances decoding accuracy, centralized training poses significant risks to user privacy, data ownership, and regulatory compliance. Especially the following three challenges remain unsolved in the existing research work: (1) lack of mechanisms for privacy-preserving feature extraction, (2) poor handling of inter-subject heterogeneity, and (3) minimal evaluation of privacy risks alongside model performance. To address these challenges, we propose FedDeepAutoCloAk a novel Federated learning framework that integrates local Deep Autoencoder-based unsupervised feature extraction and KMeans Cluster Optimization with Adaptive Knowledge for improving MI and ERP classification while minimizing information loss and strengthening privacy protection for stroke patient data. In this framework raw EEG data stay local, with cluster centroids homomorphically encrypted before secure server-side aggregation, ensuring confidentiality. The framework was evaluated on two post-stroke publicly available EEG datasets, achieved better classification performance using three deep learning models. This work provides a scalable, secure, and personalized solution for decentralized EEG-based BCIs, advancing both technical robustness and ethical integrity in neurotechnology.

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

Federated Deep Learning Framework for EEG Classification in BCI Applications

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Taslima Khanam,
  • Siuly Siuly,
  • Kate Wang,
  • Frank Whittaker,
  • Hua Wang

摘要

Electroencephalography (EEG)-based Brain – Computer Interface (BCI) systems enable direct communication between the brain and external devices, facilitating neurorehabilitation. Motor imagery (MI) classification and event-related potential (ERP) detection are two critical paradigms for developing efficient EEG-based BCIs. While deep learning enhances decoding accuracy, centralized training poses significant risks to user privacy, data ownership, and regulatory compliance. Especially the following three challenges remain unsolved in the existing research work: (1) lack of mechanisms for privacy-preserving feature extraction, (2) poor handling of inter-subject heterogeneity, and (3) minimal evaluation of privacy risks alongside model performance. To address these challenges, we propose FedDeepAutoCloAk a novel Federated learning framework that integrates local Deep Autoencoder-based unsupervised feature extraction and KMeans Cluster Optimization with Adaptive Knowledge for improving MI and ERP classification while minimizing information loss and strengthening privacy protection for stroke patient data. In this framework raw EEG data stay local, with cluster centroids homomorphically encrypted before secure server-side aggregation, ensuring confidentiality. The framework was evaluated on two post-stroke publicly available EEG datasets, achieved better classification performance using three deep learning models. This work provides a scalable, secure, and personalized solution for decentralized EEG-based BCIs, advancing both technical robustness and ethical integrity in neurotechnology.