Abstract
How can we bring both privacy and energy-efficiency to a neural system? In this paper, we propose PrivateSNN, which aims to build low-power Spiking Neural Networks (SNNs) from a pre-trained ANN model without leaking sensitive information contained in a dataset. Here, we tackle two types of leakage problems: 1) Data leakage is caused when the networks access real training data during an ANN-SNN conversion process. 2) Class leakage is caused when class-related features can be reconstructed from network parameters. In order to address the data leakage issue, we generate synthetic images from the pre-trained ANNs and convert ANNs to SNNs using the generated images. However, converted SNNs remain vulnerable to class leakage since the weight parameters have the same (or scaled) value with respect to ANN parameters. Therefore, we encrypt SNN weights by training SNNs with a temporal spike-based learning rule. Updating weight parameters with temporal data makes SNNs difficult to be interpreted in the spatial domain. We observe that the encrypted PrivateSNN eliminates data and class leakage issues with a slight performance drop (less than 2%) and significant energy-efficiency gain (about 55x) compared to the standard ANN. We conduct extensive experiments on various datasets including CIFAR10, CIFAR100, and TinyImageNet, highlighting the importance of privacy-preserving SNN training.
Introduction
Spiking Neural Networks (SNNs) are a promising solution for energy-efficient artificial intelligence, particularly on neuromorphic hardware. By utilizing discrete, binary spikes, SNNs significantly reduce power consumption compared to traditional Artificial Neural Networks (ANNs). However, the most effective way to deploy high-accuracy SNNs—converting a pre-trained ANN—introduces critical privacy vulnerabilities. PrivateSNN is a framework designed to mitigate these risks while maintaining the efficiency benefits of spiking architectures.
The Privacy Challenge in SNN Conversion
The conversion process typically suffers from two major privacy "leaks":
- Data Leakage: Standard conversion algorithms require access to the original training dataset to set neuron firing thresholds. If the training data contains sensitive information, such as medical records, sharing it for conversion is a security risk.
- Class Leakage: Even after deployment, malicious actors can use gradient-based techniques to reverse-engineer model weights and reconstruct representative images of specific classes (e.g., a specific face or object).
Methodology
The authors propose a two-step pipeline to secure the SNN conversion process:
Data-Free Conversion
Instead of using real data, PrivateSNN generates synthetic images directly from the pre-trained ANN. By starting with random noise and iteratively optimizing it to match the feature distributions in the ANN's final layers, the framework creates proxy data for threshold calibration without exposing sensitive original samples.
Weight Encryption via Temporal Learning
To prevent class reconstruction, the SNN is fine-tuned using Spatio-Temporal Back-Propagation (STBP). By training the network to process data in the temporal domain, the weights become "encrypted" relative to the spatial domain. This makes it mathematically difficult for an attacker to use standard spatial gradients to reconstruct class images.
On-Device Distillation
To ensure accuracy with synthetic data, Knowledge Distillation is used. The SNN mimics the output distribution of the teacher ANN, which acts as a regularizer during the encryption training phase.
Experimental Results
Performance and Efficiency
The researchers evaluated PrivateSNN using the VGG16 architecture on CIFAR10, CIFAR100, and TinyImageNet. The results demonstrate that privacy does not have to come at the cost of utility:
- High Efficiency: PrivateSNN achieved approximately 55x higher energy efficiency compared to a standard ANN.
- Minimal Accuracy Trade-off: The accuracy drop was less than 2% across all tested datasets compared to standard, non-private conversion methods.
- Robust Defense: Using the Frechet Inception Distance (FID) metric, the authors showed that PrivateSNN significantly degraded an attacker's ability to recover recognizable class features compared to unprotected models.
Conclusion
PrivateSNN demonstrates that the inherent temporal nature of SNNs can be leveraged as a security feature. By combining data-free conversion with temporal weight encryption, the framework provides a path for deploying secure, low-power AI on edge devices. Future work will explore scaling these techniques to more complex architectures like ResNets and Transformers.
Summary updated