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 frontier in neuromorphic computing, offering significant energy efficiency by mimicking biological neural processing through binary spikes over time. This makes them ideal for battery-constrained edge devices. However, the standard pathway to high-performing SNNs—converting pre-trained Artificial Neural Networks (ANNs)—introduces critical security vulnerabilities. PrivateSNN addresses these risks by ensuring that the conversion and deployment processes do not compromise sensitive training data.
The Privacy Challenge: Data and Class Leakage
The transition from ANNs to SNNs is fraught with two primary leakage vectors. Data Leakage occurs during the conversion process itself, as traditional methods require access to the original training dataset for calibration. If this data contains sensitive information, such as medical records, the conversion process becomes a security liability.
Class Leakage occurs after deployment. Because converted SNN weights often mirror the spatial weights of the original ANN, an attacker can reverse-engineer these parameters to reconstruct representative images of the training classes, effectively stealing the model's "knowledge" of specific categories.
The PrivateSNN Framework
To mitigate these risks, the authors propose a two-pronged defense strategy:
1. Data-Free Conversion via Synthetic Generation
Instead of using real training data, PrivateSNN generates synthetic images directly from the pre-trained ANN. By leveraging the relationship between class weights (using a Dirichlet distribution), the framework creates representative noise that mimics the original data distribution. This allows for SNN calibration without ever exposing the original sensitive samples.
2. Weight Encryption through Temporal Training
To prevent class reconstruction, PrivateSNN "encrypts" the network weights. Rather than simply copying ANN weights, the SNN is fine-tuned using Spatio-Temporal Back-Propagation (STBP). Because SNNs process information across time-steps, training them in the temporal domain ensures that the resulting weights are non-interpretable in the spatial domain, effectively scrambling the "footprint" that attackers use for reconstruction.
3. On-Device Distillation
To maintain performance while training on synthetic data, PrivateSNN employs Knowledge Distillation. The pre-trained ANN acts as a "teacher," guiding the SNN "student" to retain high accuracy and generalization capabilities despite the lack of raw training files.
Experimental Results
The framework was evaluated on CIFAR-10, CIFAR-100, and TinyImageNet benchmarks. Key findings include:
- Privacy Protection: Reconstruction attacks on PrivateSNN yielded only noisy, unrecognizable images, quantified by high Frechet Inception Distance (FID) scores compared to standard SNNs.
- Energy Efficiency: PrivateSNN demonstrated a 55× energy efficiency gain over standard VGG16 ANNs on the CIFAR-10 dataset.
- Performance: The privacy measures introduced a minimal "tax," with PrivateSNN achieving 89.2% accuracy on CIFAR-10, representing only a ~2% drop from the original ANN.
Conclusion
PrivateSNN demonstrates that privacy and energy efficiency are not mutually exclusive in neuromorphic computing. By leveraging the unique temporal properties of SNNs and utilizing data-free conversion techniques, PrivateSNN provides a robust framework for deploying secure, low-power intelligence on edge devices. Future work will focus on refining synthetic data generation and testing resilience against advanced membership inference attacks.
Summary updated