Lossless steganographic network via model arithmetic operations.

Abstract

Deep neural networks (DNN) have been effectively applied to perform steganographic tasks, demonstrating satisfactory performance. To ensure the secure execution of these tasks, it is essential to transmit the neural network in a secure and covert communication. A more covert communication method for transmission neural networks is to embed the neural network performing the secret task into the neural network performing the normal task. However, the existing methods for covert transmission of neural networks fail to completely transfer the neural network used for performing the secret task. This will affect the normal use of the neural network. To address this issue, this paper proposes lossless steganographic network via model arithmetic operations, which ensures that the performance and integrity of the neural network performing the secret task are not compromised ordamaged during the steganography process. We hide the parameters of the secret model using arithmetic operations between the parameters of stego model, and then train the stego model based on this to achieve the purpose of hiding the network. To ensure the stego model can effectively perform the steganography task, we employ an iterative training approach. During each training iteration, new parameters are computed and subsequently updated. Experiments show that this steganographic transmission method for secret model can securely and losslessly deliver high-performing steganographic networks to recipients.