Title: STGN: spatio-temporal graph network for few-shot cross-domain image steganalysis

Authors: Mingqian Liu; Daqiu Li; Xiang Zhang; Zhangjie Fu

Addresses: Engineering Research Center of Digital Forensics, Ministry of Education, Nanjing University of Information Science and Technology, Nanjing, 210044, China ' Nanjing Vocational College of Information and Technology, Nanjing, 210023, China ' Engineering Research Center of Digital Forensics, Ministry of Education, Nanjing University of Information Science and Technology, Nanjing, 210044, China ' Engineering Research Center of Digital Forensics, Ministry of Education, Nanjing University of Information Science and Technology, Nanjing, 210044, China

Abstract: In the actual image steganalysis task, it is difficult for steganalysis models to obtain large-scale steganographic images of unknown steganography as training datasets. Inspired by few-shot learning, we propose a novel spatio-temporal graph network (STGN) for few-shot cross-domain steganalysis. Firstly, we design a multi-domain feature preprocessing network in the spatial, frequency, and feature domains, so that STGN can extract the deep steganographic features from different domains. Secondly, we design a spatiotemporal graph convolution network is designed to extract the effective spatio-temporal components in the steganalysis sequence feature; Then, the spatio-temporal steganalysis feature is as the input of graph network to classify. Finally, the STGN is trained on the benchmark datasets including the steganographic images of spatial and frequency domains. Through intra-domain and cross-domain testing, experimental results show that the average accuracy is above 85.55% (1-shot) and 93.97%(5-shot).

Keywords: image steganalysis; few-shot learning; multi-domain feature extraction; graph convolution network; cross-domain steganalysis; deep learning.

DOI: 10.1504/IJAACS.2025.148551

International Journal of Autonomous and Adaptive Communications Systems, 2025 Vol.18 No.4, pp.324 - 340

Received: 12 Mar 2024
Accepted: 23 May 2024
Published online: 11 Sep 2025 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article