TL;DR
This paper presents the first comprehensive evaluation of Membership Inference Attacks (MIAs) on deep Spiking Neural Networks (SNNs). The authors demonstrate that while SNNs are highly energy-efficient, they are often more vulnerable to privacy leaks than traditional Artificial Neural Networks (ANNs) when trained on neuromorphic data. However, the authors also show that converting trained ANNs to SNNs can serve as an effective passive defense against these privacy threats.
What Problem Does This Paper Address?
Spiking Neural Networks (SNNs) are often hailed as the third generation of neural networks. Unlike traditional Artificial Neural Networks (ANNs) that process continuous floating-point values, SNNs mimic biological brains by processing discrete binary "spikes" over time. This temporal, event-driven processing makes them incredibly energy-efficient and ideal for edge computing and neuromorphic hardware.
However, as SNNs transition into real-world applications like healthcare and autonomous driving, their privacy risks remain largely unexplored. In traditional deep learning, Membership Inference Attacks (MIAs)—where an adversary determines whether a specific data point was used to train a target model—pose a severe privacy threat. This paper tackles a fundamental question: Does the unique temporal and spiking nature of SNNs protect user privacy, or does it introduce new vulnerabilities?
Methods
To evaluate SNN privacy, the authors adapted seven established ANN-based MIAs and introduced one entirely new SNN-specific attack strategy.
Because SNNs output binary spikes rather than continuous probability scores (confidences), standard ANN attacks cannot be applied directly. To bridge this gap, the authors leveraged the fire rate of spiking neurons as a proxy for prediction confidence, since SNN training aligns fire rates with one-hot labels via Mean Squared Error (MSE) loss. They also introduced a novel attack feature: the average membrane potential of the network's final layer.
The attacks were evaluated against deep SNN architectures (CNN, VGG11, ResNet18) trained via two primary paradigms:
- Backpropagation-based supervised learning: SNNs trained directly using surrogate gradients (specifically the ArcTangent or Atan function) to bypass the non-differentiable nature of the spiking threshold.
- ANN-to-SNN Conversion: Taking a standard pre-trained ANN and converting it into an SNN by swapping ReLU layers for spiking neurons and applying weight normalization.
Finally, the authors evaluated data augmentation techniques—including EventDrop and Neuromorphic Data Augmentation (NDA)—as potential privacy defenses.
Datasets
The authors evaluated their methods using three neuromorphic (event-based) datasets alongside their static counterparts to ensure a fair comparison:
- N-MNIST vs. standard MNIST
- CIFAR10-DVS vs. standard CIFAR-10
- N-Caltech101 vs. standard Caltech101
Key Insights and Ideas
- The Neuromorphic Privacy Penalty: SNNs trained directly on neuromorphic data exhibit a significantly larger generalization gap (the difference between training and testing accuracy) than ANNs. This wider gap makes them highly susceptible to MIAs.
- Conversion as a Shield: Converting a pre-trained ANN into an SNN acts as an effective passive defense. Because the converted SNN utilizes weights differently through spiking dynamics rather than continuous activations, it naturally obfuscates some of the membership information, lowering attack success rates.
- Generalization Gap Rules SNNs Too: Just as in traditional deep learning, the generalization gap is the strongest indicator of MIA vulnerability in SNNs. Hyperparameters that increase this gap—such as using Leaky Integrate-and-Fire (LIF) neurons combined with an Atan surrogate function—directly amplify privacy risks.
Key Results
- Higher Vulnerability on Neuromorphic Data: SNNs trained on neuromorphic datasets showed up to a 10% higher balanced attack accuracy compared to equivalent ANNs.
- Mitigation via Conversion: Converting ANNs to SNNs dropped the MIA accuracy by up to 11.5%, though this came at the cost of a 7.6% reduction in the model's primary classification accuracy.
- Augmentation Defenses: Neuromorphic Data Augmentation (NDA) proved superior to EventDrop, slashing MIA performance by up to 25.7%. However, it could not completely eliminate the threat, leaving attack accuracies hovering between 51.7% and 66.4% (where 50% represents random guessing).
Limitations and Future Work
A major drawback of using data augmentation as a defense is the associated utility loss; target model test accuracy dropped by 10% to 20% on static datasets when basic augmentations were applied.
The authors note that future research must focus on finding novel ways to shrink the generalization gap in SNNs without compromising their training performance or edge-computing efficiency. Additionally, developing tailored cryptographic or Differential Privacy (DP) frameworks specifically optimized for spiking architectures remains an open and urgent challenge.
Summary updated