Istego100k++: A Large-Scale GAN-Based Steganalysis Dataset Using Adversarially Learned Embedding Costs

Abstract

Steganalysis benchmarks built with fixed distortion functions — such as nsF5, J-UNIWARD, and UERD — are increasingly inadequate against GAN-based steganography, where adversarially learned embedding costs bypass traditional rich-model detectors. To address this gap, we introduce IStego100K++, a large-scale GAN-generated JPEG stego dataset of 100,071 cover–stego image pairs at 1024×1024 resolution. Stego images were produced using a pre-trained UT-GAN generator with a U-Net architecture and a differentiable Double-Tanh embedding simulator, covering payload rates of {0.1–0.5} bpp and JPEG quality factors of {75–95}, achieving a mean PSNR of 74.52 dB, a mean SSIM of 0.9994, and 37.9% of images recording zero measurable distortion after JPEG recompression. Evaluated against two established rich-model steganalyzers — DCTR and GFR — detection accuracy drops from 79.55% on the original IStego100K to approximately 50% across all tested payload rates and quality conditions. Additional evaluation with CNN-based steganalyzers (SRNet-KV and XuNet) yields AUC values of 0.6953 and 0.6901 respectively, confirming that while CNN detectors outperform rich-model methods, IStego100K++ remains a significantly more challenging benchmark for next-generation universal steganalysis research.