AI Countermeasures in Steganography

The rise of AI-based steganalysis presents a significant challenge to modern steganographic technologies. Convolutional neural networks (CNNs) excel in detecting traditional methods like WOW, HILL, and S-UNIWARD, with accuracy rates exceeding 90% at high payloads. This chapter examines the resilience of Spread Spectrum Image Steganography (SSIS) against AI detectors, using the SRNet architecture on the BOWS2 dataset. Results show that at a payload of 0.125 bpp, SSIS achieves a detection error rate of 0.397, significantly outperforming classical methods. The chapter highlights how the global distribution of embedded information disrupts CNNs and complicates feature extraction. Fine-tuning strategies reveal the potential for AI detectors to specialize in SSIS detection, improving accuracy from 60.30% to 72.15%, albeit with a trade-off in performance on traditional methods. Recommendations include parameter selection and adaptive strategies for the SSIS application. Overall, the findings contribute to understanding the ongoing dynamics between steganography and steganalysis in the AI era.