Characterizing Privacy Leakage in Spiking Neural Networks: A Systematic Analysis of Temporal Dynamics, Stochasticity, Neuronal Dynamics and Learning Mechanisms

Ayana Moshruba, Maryam ParsaarXiv 2026 · arXiv

TL;DR

Spiking Neural Networks (SNNs) aren't inherently private just because they use discrete spikes; rather, their privacy is a complex byproduct of how information is organized in time. This paper systematically disentangles these factors, revealing that while burst-like firing patterns can hide data, the real "privacy-utility win" comes from using biologically expressive adaptive neurons (AdEx), which protect data without the typical performance penalties.

What Problem Does This Paper Address?

As machine learning moves to the edge—in healthcare, autonomous systems, and wearables—privacy becomes a critical bottleneck. Spiking Neural Networks (SNNs) have emerged as a promising, energy-efficient alternative to traditional Artificial Neural Networks (ANNs). Early research suggested SNNs might be naturally more resistant to Membership Inference Attacks (MIAs)—where an adversary tries to determine if a specific data point was used to train a model—but the results were inconsistent.

The core challenge is that SNNs are "black boxes" of temporal complexity. When you change how a network fires, you simultaneously change its randomness (stochasticity), the exact timing of its spikes, and the way it learns. Until now, we didn't know which of these levers actually drives privacy. The authors aim to isolate these variables to provide a principled basis for designing privacy-preserving neuromorphic systems.

Methods

The authors employ a "divide and conquer" experimental framework to isolate four key components:

  1. Temporal Representation: They compare standard Rate-based coding (intensity = frequency) against biologically inspired Poisson-Burst and Delayed-Burst patterns.
  2. Stochasticity (Noise): They inject controlled randomness at three levels: temporal jitter (shifting spike times), input noise (perturbing the signal before spikes), and spike noise (randomly adding/deleting spikes).
  3. Neuronal Dynamics: They compare the standard Leaky Integrate-and-Fire (LIF) neuron against the more complex Adaptive Exponential Integrate-and-Fire (AdEx) model, which has internal variables for adaptation.
  4. Learning Mechanisms: They evaluate Backpropagation Through Time (BPTT) using surrogate gradients versus Evolutionary Learning (EONS), which optimizes networks without gradient information.

Privacy is measured using a standard black-box shadow-model MIA, where an attack classifier tries to distinguish between "member" (training) and "non-member" (test) samples based on the target model's outputs.

Datasets

The study evaluates these dynamics across five diverse benchmarks:

  • Image Classification: MNIST, Fashion-MNIST (FMNIST), and CIFAR-10.
  • Tabular Data: Iris and Breast Cancer Wisconsin (Diagnostic).

Key Insights and Ideas

  • Structure Over Noise: One of the most surprising findings is that simply adding noise (stochasticity) is a poor privacy strategy. While noise degrades accuracy, it doesn't consistently reduce leakage. Instead, the temporal arrangement of spikes—how they are clustered or delayed—is what actually thwarts the attacker.
  • The AdEx "Free Lunch": In many privacy-preserving techniques, there is a strict trade-off: more privacy equals less accuracy. However, AdEx neurons seem to break this rule. By generating complex firing patterns internally, AdEx neurons reduced membership leakage while improving or maintaining accuracy compared to simpler LIF neurons.
  • Learning Sets the Floor: The choice of optimization algorithm determines the baseline privacy. Evolutionary learning (EONS) resulted in MIA performance near 0.5 (random guessing), meaning the models were inherently private. Gradient-based learning (BPTT) leaked significantly more, though temporal dynamics could be used to mitigate this.
  • Heterogeneity Matters: Mixing different firing patterns (e.g., 80% Rate, 10% Burst, 10% Delayed) can reduce leakage without the massive accuracy drops seen in purely burst-based networks.

Key Results

  • Bursting Benefits: On F-MNIST, switching from Rate to Poisson-Burst reduced the MIA AUC (Area Under the Curve) from 0.534 to 0.502 (near-perfect privacy), though it often came with an accuracy cost.
  • AdEx Superiority: Across all datasets, AdEx neurons consistently produced lower MIA AUCs than LIF neurons. On CIFAR-10, AdEx recovered a substantial portion of the utility lost when using burst dynamics under the LIF model.
  • Evolutionary Privacy: For tabular datasets, evolutionary training achieved an MIA AUC of ~0.5 regardless of the firing pattern, suggesting that the lack of direct gradient-to-sample coupling is a powerful privacy feature.
  • DPSGD Interaction: When using Differentially Private SGD (DPSGD), the authors found that Delayed-Burst dynamics were highly sensitive to strict privacy budgets but recovered quickly as the budget increased, whereas Poisson-Burst suffered a persistent "utility deficit."

Limitations and Future Work

  • Attack Scope: The study focuses exclusively on black-box MIAs. The authors acknowledge that white-box attacks or model inversion attacks might reveal different vulnerabilities.
  • Computational Cost: Due to the high cost of evolutionary optimization, those experiments were limited to smaller tabular datasets. Scaling this analysis to larger models like those needed for CIFAR-10 remains a challenge.
  • Internal Leakage: The paper looks at the final output of the network. A promising next step is "tracking the leakage" through successive layers and time steps to see exactly where the training data signal begins to fade.

Summary updated

Abstract

Spiking neural networks (SNNs) encode information not only in whether neurons fire but also in when they fire, and this temporal dimension shapes what a trained network reveals about its training data. Biologically inspired firing patterns can reduce membership leakage in SNNs without changing the network architecture, which positions temporal dynamics as a design variable for privacy. A change in firing behavior, however, alters stochasticity, spike timing, neuronal response, and the learning problem at once, so harnessing this effect requires identifying which of these properties produces it. This study traces that origin by separating temporal representation, stochasticity, intrinsic neuron dynamics, and learning within a common experimental framework, evaluated under a black-box membership inference attack across image and tabular benchmarks. Varying the temporal representation shows that burst-like firing patterns reduce leakage relative to regular spiking, but at a cost in accuracy. Injecting randomness into the input, spike timing, and spike trains does not reproduce this reduction in leakage, which shows that the reduction arises from how spikes are arranged in time rather than from noise. The privacy benefit further depends on where these firing patterns originate. When neurons generate these patterns through their own adaptive dynamics, leakage falls while accuracy is preserved, so privacy no longer comes at the expense of performance. The learning algorithm then sets the baseline leakage. Under evolutionary training, the attack already performs near chance level, leaving temporal dynamics little room to improve privacy further. Privacy in SNNs is therefore not an inherent property of spikes but a product of how temporal activity is generated, processed, and learned, which offers a principled basis for designing neuromorphic systems that protect sensitive data.
BibTeX
@misc{moshruba2026characterizing,
  author = {Ayana Moshruba and Maryam Parsa},
  title = {{Characterizing Privacy Leakage in Spiking Neural Networks: A Systematic Analysis of Temporal Dynamics, Stochasticity, Neuronal Dynamics and Learning Mechanisms}},
  year = {2026},
  eprint = {2610.36243},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2610.36243},
}
Figure