On the Sensitivity of Firing Rate-Based Federated Spiking Neural Networks to Differential Privacy
Luiz Pereira, Mirko Perkusich, Dalton Valadares, Kyller Gorgônio
TL;DR
This work analyzes how differential privacy via DP-SGD perturbs firing-rate signals in LIF-based Spiking Neural Networks used in Federated SNN Learning. By deriving a first-order sensitivity analysis, the authors connect clipping-induced bias and noise-induced variance in DP-SGD to shifts in firing-rate statistics, and they map these perturbations to downstream rate-dependent coordination decisions such as rate-weighted aggregation and rate-difference client selection under non-IID data. Ablation on Google Speech Commands demonstrates systematic rate shifts, attenuated aggregation, and unstable client rankings under privacy constraints, with rate shifts correlating to sparsity and memory indicators. The findings offer actionable guidance for privacy-preserving FNL, highlighting a privacy-utility trade-off and suggesting calibration strategies for rate-aware coordination under DP.
Abstract
Federated Neuromorphic Learning (FNL) enables energy-efficient and privacy-preserving learning on devices without centralizing data. However, real-world deployments require additional privacy mechanisms that can significantly alter training signals. This paper analyzes how Differential Privacy (DP) mechanisms, specifically gradient clipping and noise injection, perturb firing-rate statistics in Spiking Neural Networks (SNNs) and how these perturbations are propagated to rate-based FNL coordination. On a speech recognition task under non-IID settings, ablations across privacy budgets and clipping bounds reveal systematic rate shifts, attenuated aggregation, and ranking instability during client selection. Moreover, we relate these shifts to sparsity and memory indicators. Our findings provide actionable guidance for privacy-preserving FNL, specifically regarding the balance between privacy strength and rate-dependent coordination.