Privacy-Preserving Federated Autoencoder for ECG Anomaly Detection on Edge Devices

The article presents a novel privacy-preserving federated autoencoder system for unsupervised ECG anomaly detection, designed for deployment on edge devices while ensuring compliance with privacy regulations like GDPR and HIPAA. The system integrates three autoencoder architectures (VanillaAE, ConvAE, VAE) with federated learning via Flower's FedAvg across ten simulated hospitals, utilizing client-side differentially private SGD (DP-SGD) and 8-bit integer post-training quantization, achieving an AUROC of 0.782. This approach enables real-time inference on constrained hardware, demonstrating that differential privacy and quantization can be effectively balanced without compromising detection performance, which is crucial for practitioners focusing on secure and efficient health monitoring solutions.