Homomorphic Encryption for Spiking Neural Networks

Nowadays, Machine learning is employed in many different everyday problems, especially in the form of Neural Networks. However, the raising computational load and the heavy resource requirements of these structures have led to solutions based on cloud-based services. Moreover, to improve the efficiency and the power consumption of the models’ calculations, it has been introduced a new generation of neural networks called Spiking Neural Network which aims to overcome the previous limitations by imitating the human brain behavior. Its sparse, dynamic and eventdriven analytic capabilities improve meaningfully the performances reducing overall costs. In the last few years, however, what has been considered concerning about this type of service is related to the privacy preservation of the confidential data evaluated by the networks held by servers. To solve the issue several strategies have been proposed, but one of the most promising is Homomorphic Encryption, a particular scheme based on a special method to encrypt information that allows some kinds of computations directly on the encrypted data without decrypting it. Performing this type of Encryption can overcome privacy concerns while using Machine Learning in cloud platforms letting the client employ a network model owned by another party by sending the encrypted data in complete safety. The project is about the application of this security strategy on 2 different generations of neural networks, the artificial and the spiking ones. Hence, the main features of the neural networks have been collected and the homomorphic encryption strategy chosen, the Brakerski/Fan-Vercauteren, has been discussed as background research. Then it has been selected a convolutional architecture, used as a model to build a spiking version of the same structure, to perform the classification of images from MNIST, FashionMNIST, and CIFAR10 datasets. After the training phases, the performances of the 2 obtained networks have been compared when a hypothetical client requires encrypted computations. The simulations have been fulfilled by means of PyTorch useful to deal with tensors and networks, the libraries NORSE and Pyfhel. The first provides the spiking functions while the latter, based especially on SEAL, has been used to deal with the encryption.