Evaluation Dataset and Analysis Scripts for Comparative Evaluation of Sequential and CNN-Seeded Randomized Embedding for Authenticated Image Steganography

We present a steganographic system that combines CNN-based, seed generation for randomized LSB embedding with AES-256-GCM authenticated encryption and Argon2id key derivation. Unlike prior work that treats randomized embedding as inherently more resistant to statistical steganalysis, we empirically evaluate this assumption rather than presume it. Across 18 test images at three embedding rates (10%, 50%, and 90% of capacity), our method preserves image quality equivalent to a sequential-LSB baseline (PSNR and SSIM values within measurement noise of one another) and achieves 100% message recovery. Chi-square analysis shows negligible difference between embedding strategies, consistent with its position-blind construction. Contrary to our initial hypothesis, RS analysis consistently shows randomized embedding as more detectable than sequential embedding across all tested rates, a result we argue is consistent with RS analysis's documented sensitivity to non-sequential embedding rather than an anomaly of our implementation. We report this finding directly rather than adjusting our claims to match expectations, and discuss its implications for the design of content-adaptive steganographic systems.