AI-based image steganalysis under limited computational resources

Abstract

This study addresses the challenges of modern steganalysis, which lies in the dichotomy between highly effective yet computationally expensive State-of-the-Art (SOTA) artificial intelligence models and lightweight architectures that are fast but incapable of independently detecting weak steganographic signals. The hypothesis proposed in this research suggests that combining classical feature engineering techniques – particularly the use of Spatial Rich Model (SRM) filters to enhance residual noise – with a modern self-supervised learning (SSL) approach for regularization and improved generalization capability can endow a lightweight convolutional neural network with the necessary properties for effective performance. To verify this hypothesis, a comprehensive comparative experiment was conducted involving four models: a baseline lightweight architecture, a model employing SRM filters, a heavy SOTA SRNet (Residual Network) model, and the proposed hybrid model. The experiment was carried out on a complex heterogeneous dataset comprising images processed by three distinct steganographic algorithms with two embedding rates. Performance evaluation was conducted on two datasets: a test sample from the same data domain (in-distribution) and a completely new, external dataset to assess generalization capability (out-of-distribution). The experimental results fully confirmed the main hypothesis. The hybrid model achieved the highest detection accuracy among lightweight approaches (AUC – Area Under the ROC Curve of 0.636) and, most importantly, demonstrated the greatest robustness to domain shift (AUC of 0.539 on the external dataset), showing the smallest degradation in performance. The study also revealed a counterintuitive effect: the heavy SOTA SRNet architecture exhibited a significant failure (AUC of 0.348) under heterogeneous data conditions, indicating its tendency to overfit to specific artifacts.