Evaluation of Activation Functions in Convolutional Neural Networks for Image Classification Based on Homomorphic Encryption

In the dynamic environment of big data and cloud computing, image feature classification has become a key factor spanning various fields. Ensuring the security, privacy, and computational efficiency of image data, while minimizing the processing of image data and maintaining the effectiveness of encrypted classification, is a significant challenge. In this paper, we propose a new method, Homomorphic Encryption Image Classification Evaluation (HEICE), for secure image classification. This method leverages the power of Convolutional Neural Networks (CNNs) and the security of Homomorphic Encryption (HE) to perform image classification on encrypted data. Each model uses different activation functions: square function, polynomial approximation of ReLU, polynomial approximation of Sigmoid and Tanh, and a piecewise linear approximation. These modified models are then used to test encrypted images, and the results are compared with the baseline. This method allows us to evaluate the performance of different activation functions when processing encrypted data and to choose the most suitable model for image classification, i.e., the classification model with the square function as the activation function. Our method provides a systematic approach to address the challenge of ensuring model performance while maintaining data security in image classification. This comparison validates the effectiveness of our method in achieving the dual objectives of maintaining data privacy and achieving accurate image classification.