Intelligent and Secure Image Data Hiding Using AI-Driven Embedding (TA-DPCNN) and Cryptographic Protection

This paper introduces a Texture-Aware Dual-Path Convolutional Neural Network (TA-DPCNN) for intelligent and secure image steganography. The dual-branch CNN jointly learns texture density and distortion sensitivity to produce a dynamic embedding strength map, enabling context-aware modulation of pixel-level data hiding while preserving visual fidelity. A hybrid AES-RSA-XOR encryption scheme and an HMAC-SHA-256 verification layer ensure confidentiality and integrity, while Hamming and Reed-Solomon codes improve robustness under noisy transmission. Experimental results on BOSSBase, ALASKA2, and UCID datasets show that TA-DPCNN achieves up to 42.8 dB PSNR and 0.97 SSIM, outperforming recent adaptive models such as HiNet and DeepSteg by$8-12 \%$in both quality and resilience. Visual residual analysis confirms adaptive embedding in high-entropy regions, minimizing perceptual distortion. TA-DPCNN thus establishes a scalable, verifiable framework for high-capacity, AI-driven steganography.