Federated Spiking GANs with Rényi Differential Privacy for privacy-preserving synthetic EEG generation in brain-computer interfaces.
- 1Department of Computer Science, Lakehead University, Thunder Bay, ON, Canada.
Abstract
Electroencephalography (EEG) data used in brain-computer and brain-machine interfaces (BCI/BMI) contain sensitive, user-specific information, creating privacy risks in consumer applications such as gaming, human augmentation, and virtual reality. Existing methods like encryption and anonymization either fail to protect data during analysis or still expose identifiable attributes. While differential privacy (DP) offers formal protection, it often reduces data utility. The key challenge is to generate synthetic EEG data that protects sensitive attributes while retaining temporal features for downstream use. To address this, we developed a federated spiking generative framework that integrates a spiking neural network (SNN)-based generator, temporally correlated noise, an artificial neural network (ANN)-based discriminator, and Rényi differential privacy (RDP) within a federated learning framework. The SNN-based generator captures temporal characteristics, and the ANN-based discriminator maintains classification accuracy. Federated training prevents centralization of raw client data, while privacy is preserved through controlled perturbation of spike-train and membrane-potential representations, supported by RDP-based privacy accounting. We evaluated our architecture against methods such as RDP-TimeGAN, RDP-CRNN-GAN, RDP-RCGAN, and RDP-Clare-GAN, demonstrating that our approach preserves temporal features and downstream utility while maintaining privacy in synthetic EEG generation. This solution enables privacy-preserving generation of distributed synthetic EEG data for BCI/BMI research and development.
Related
- FedEdgeR: federated and privacy-preserving edgeR for differential gene expression analysis
- Regulating Next-Generation Implantable Brain-Computer Interfaces: Recommendations for Ethical Development and Implementation
- MediNet: simplifying federated and privacy-preserving AI deployment in healthcare
- BiNVaST: niche boundary-preserving and synthetic data-guided variational autoencoders for concordant inference across high-resolution spatial transcriptomic platforms
- Evaluating in-context learning with prompting regimes for transformer-based synthetic health tabular data generation
BibTeX
@article{paul2026federated,
title = {{Federated Spiking GANs with Rényi Differential Privacy for privacy-preserving synthetic EEG generation in brain-computer interfaces.}},
author = {Shouvik Paul and Garima Bajwa},
journal = {Frontiers in human neuroscience},
year = {2026},
month = oct,
doi = {10.3389/fnhum.2026.1909432},
url = {https://europepmc.org/article/MED/42952986},
note = {CC BY},
}
RIS
TY - JOUR
TI - Federated Spiking GANs with Rényi Differential Privacy for privacy-preserving synthetic EEG generation in brain-computer interfaces.
AU - Shouvik Paul
AU - Garima Bajwa
PY - 2026
DA - 2026/10/14
JO - Frontiers in human neuroscience
DO - 10.3389/fnhum.2026.1909432
UR - https://europepmc.org/article/MED/42952986
AB - Electroencephalography (EEG) data used in brain-computer and brain-machine interfaces (BCI/BMI) contain sensitive, user-specific information, creating privacy risks in consumer applications such as gaming, human augmentation, and virtual reality. Existing methods like encryption and anonymization either fail to protect data during analysis or still expose identifiable attributes. While differential privacy (DP) offers formal protection, it often reduces data utility. The key challenge is to generate synthetic EEG data that protects sensitive attributes while retaining temporal features for downstream use. To address this, we developed a federated spiking generative framework that integrates a spiking neural network (SNN)-based generator, temporally correlated noise, an artificial neural network (ANN)-based discriminator, and Rényi differential privacy (RDP) within a federated learning framework. The SNN-based generator captures temporal characteristics, and the ANN-based discriminator maintains classification accuracy. Federated training prevents centralization of raw client data, while privacy is preserved through controlled perturbation of spike-train and membrane-potential representations, supported by RDP-based privacy accounting. We evaluated our architecture against methods such as RDP-TimeGAN, RDP-CRNN-GAN, RDP-RCGAN, and RDP-Clare-GAN, demonstrating that our approach preserves temporal features and downstream utility while maintaining privacy in synthetic EEG generation. This solution enables privacy-preserving generation of distributed synthetic EEG data for BCI/BMI research and development.
SN - 1662-5161
ER -