BST-NS: A Biologically Inspired Spectral-Temporal Noise Shaping Framework for Privacy-Preserving Federated ECG Learning

Abstract

Federated learning enables collaborative model training across distributed electrocardiogram (ECG) datasets while preserving patient privacy, but conventional differential privacy mechanisms degrade model accuracy by adding isotropic Gaussian noise without regard for the clinical structure of ECG signals. We propose a biologically inspired spectral-temporal noise shaper (BST-NS) that replaces standard noise injection with an adaptive, frequency-selective perturbation process tailored to preserve diagnostically relevant ECG features. Our method decomposes client-side gradient updates using empirical mode decomposition, then applies a fractional-order filter who’s cutoff and roll-off are dynamically controlled by a recurrent neural network. This network processes spectral power metrics and heart rate variability statistics from the local ECG data to output two parameters: a fractional filter order and a noise amplitude scaling factor. The resulting noise vector concentrates its energy in frequency bands outside the clinically critical range of 0.04 to 40 hertz, where cardiac waveform morphology and arrhythmia signatures reside. We realize the filter using the Oustaloup approximation and perform noise generation via frequency-domain convolution for computational efficiency. In our federated framework, each wearable client adds this structured noise to its clipped gradient before sending the update to a cloud aggregator, which then performs global model averaging. The privacy guarantee is verified using the moments accountant to achieve differential privacy with epsilon equal to one and delta equal to ten to the minus five. Our primary contribution is a privacy-preserving mechanism that decouples utility from noise magnitude by exploiting the frequency-domain characteristics of ECG gradients. This approach significantly reduces distortion of PQRST waveform morphology and heart rate variability metrics compared to isotropic noise injection, thereby maintaining high arrhythmia classification accuracy. The work demonstrates that biologically motivated noise shaping can reconcile strong privacy protections with clinically meaningful model performance in cloud-based health data aggregation systems.