SpyKing-Privacy-preserving framework for Spiking Neural Networks

Area of Science:

• Computer Science
• Artificial Intelligence
• Cryptography

Background:

• Deep neural networks (DNNs) are widely used but process sensitive data, raising privacy concerns.
• Fully Homomorphic Encryption (FHE) allows computations on encrypted data, offering a solution for privacy-preserving AI.
• Spiking Neural Networks (SNNs) mimic biological neurons and may offer advantages in encrypted computations.

Purpose of the Study:

• To evaluate the performance of FHE applied to DNNs.
• To compare FHE-applied DNNs with SNNs for privacy-preserving AI.
• To analyze challenges in encrypted neural computations, especially non-linear operations.

Main Methods:

• Experiments were conducted on MNIST, FashionMNIST, and CIFAR10 datasets.
• The SpyKing framework was used to analyze encrypted neural computations.
• Encryption parameters were systematically varied to optimize SNN performance.

Main Results:

• FHE significantly increases computational costs but maintains accuracy and data security.
• SNNs achieved up to 35% higher absolute accuracy than DNNs on encrypted data with low plaintext modulus values.
• Limitations of FHE in handling non-linear operations were identified.

Conclusions:

• FHE is a viable, albeit computationally intensive, technology for privacy-preserving AI.
• SNNs demonstrate significant potential for enhancing accuracy in privacy-preserving AI applications.
• There is a growing need for secure and efficient neural computing solutions.