DeepStegAI: A High-Density Multi-File Steganography Architecture with Adaptive Visual Masking and XAI-Driven Steganalysis

Abstract

This paper presents DeepStegAI, a full-stack steganographic architecture that integrates cryptographic protocols, adaptive spatial embedding, and deep convolutional steganalysis into a unified, deployable pipeline. The primary contribution is the Adaptive Edge module, which derives a boundary map using the Canny operator applied to the mostsignificant bits (MSB) of the green channel of colour cover images. Gradient thresholds in the range 100-200 identify edge regions where the Human Visual System (HVS) exhibits reduced sensitivity, enabling variable payload allocation of up to 3 bits per channel (BPC) at edges and 1 BPC in smooth regions—yielding a mean embedding density of 1.58 bpp. For greyscale images, the luminance plane serves as the single available channel, maintaining methodological consistency. All concealed data undergoes pre-encryption with AES-256 (Fernet) coupled with PBKDF2-HMAC-SHA256 key stretching at 480,000 iterations. Steganalysis is performed by StegoCNN, a lightweight convolutional evaluator employing Spatial Rich Model (SRM) kernels. Experiments were conducted on the BOSSBase 1.01 benchmark dataset augmented to 4,000 training tensors. Based on internal evaluation, the framework achieves a PSNR of 62.1 dB, an SSIM of 0.9984, and a steganalysis accuracy of 99.98% on the held-out test partition. These results indicate competitive performance relative to published edge-based and transformer-based methods; broader generalisation across diverse image domains warrants further investigation.