Towards Deep Network Steganography: From Networks to Networks

TL;DR

This paper introduces a novel deep network steganography method that covertly transmits DNN models by disguising secret-learning tasks as ordinary tasks through gradient-based filter insertion and side information hiding.

Contribution

It proposes a learning task-oriented steganography scheme for DNNs using gradient-based filter insertion and key-based embedding, enabling covert model transmission.

Findings

Effective in intra-task steganography

Effective in inter-task steganography

Demonstrates robustness and concealment

Abstract

With the widespread applications of the deep neural network (DNN), how to covertly transmit the DNN models in public channels brings us the attention, especially for those trained for secret-learning tasks. In this paper, we propose deep network steganography for the covert communication of DNN models. Unlike the existing steganography schemes which focus on the subtle modification of the cover data to accommodate the secrets, our scheme is learning task oriented, where the learning task of the secret DNN model (termed as secret-learning task) is disguised into another ordinary learning task conducted in a stego DNN model (termed as stego-learning task). To this end, we propose a gradient-based filter insertion scheme to insert interference filters into the important positions in the secret DNN model to form a stego DNN model. These positions are then embedded into the stego DNN model using a key by side information hiding. Finally, we activate the interference filters by a partial optimization strategy, such that the generated stego DNN model works on the stego-learning task. We conduct the experiments on both the intra-task steganography and inter-task steganography (i.e., the secret and stego-learning tasks belong to the same and different categories), both of which demonstrate the effectiveness of our proposed method for covert communication of DNN models.