Elastic Supernet with Dynamic Training for JPEG steganalysis

Abstract

JPEG is the predominant image format across social networks, serving as a prime cover medium for image steganography. However, previous deep learning models for JPEG steganalysis heavily rely on domain expertise and tedious trial-and-error methods. In this paper, we propose a two-stage neural architecture search scheme for JPEG steganalysis, based on Elastic Supernet with Dynamic Training (ESDT). The method involves constructing a weight-nesting supernet, with the largest subnetwork pretrained on ImageNet (a large-scale visual database widely used for pretraining deep learning models) and finetuning for JPEG steganalysis. Based on this pretrained network, we aim to enhance the model’s performance in downstream tasks while reducing reliance on domain knowledge. A progressive shrinking strategy is introduced during supernet training to accommodate the need of elastic kernel sizes, depths, and widths. In the final stage, we utilize a performance predictor to identify the optimal subnetwork within the refined supernet. Extensive experiments showcase the method’s superiority over state-of-the-art methods in JPEG steganalysis, achieving lower computational costs and superior generalization performance.