Convolutional neural network model over encrypted data based on functional encryption

Currently, homomorphic encryption, secure multi-party computation, and other encryption schemes are used to protect the privacy of sensitive data in outsourced convolutional neural network (CNN) models. However, the computational and communication overhead caused by the above schemes would reduce system efficiency. Based on the low cost of functional encryption, a new convolutional neural network model over encrypted data was constructed using functional encryption. Firstly, two algorithms based on functional encryption were designed, including inner product functional encryption and basic operation functional encryption algorithms to implement basic operations such as inner product, multiplication, and subtraction over encrypted data, reducing computational and communication costs. Secondly, a secure convolutional computation protocol and a secure loss optimization protocol were designed for each of these basic operations, which achieved ciphertext forward propagation in the convolutional layer and ciphertext backward propagation in the output layer. Finally, a secure training and classification method for the model was provided by the above secure protocols in a module-composable way, which could simultaneously protect the confidentiality of user data as well as data labels. Theoretical analysis and experimental results indicate that the proposed model can achieve CNN training and classification over encrypted data while ensuring accuracy and security. © 2024 Editorial Board of Journal on Communications. All rights reserved.