A Lightweight Deep Learning Framework for Secure Image Steganography in Sketch Domains
DOI:
https://doi.org/10.51483/IJAIML.6.9s.2026.1774-1787Keywords:
Image Steganography, Deep Neural Networks, Adversarial Training, Sketch Image Embedding, Real-Time Encoding, Runtime Analysis, Visual Performance Evaluation.Abstract
In the era of the digital age, when a single pixel could be hiding confidential information, the necessity of secure, real-time steganography is more relevant than ever. The current paper introduces SteganoNet++, a deep-learning system designed to hide hidden image information within sketch-like images while keeping high quality and detail. It can be described as a lightthe authorsight encoder-decoder system, strengthened with adversarial training, which can demonstrate solid performance in even sparse visual realms.
One such central innovation is a dynamic generator, the SketchDataset, which provides different secret-cover image pairs on the fly, making the model a better generalizer. Experimental results reveal notably high reconstruction performance (SSIM = 0.9865, PSNR = 36.03 dB), though stego images still exhibit visible artifacts (SSIM = –0.2875, PSNR = –5.84 dB). Nevertheless, the framework is strong when data fidelity is the primary value rather than total invisibility, as in digital watermarking and forensic surveillance. SteganoNet++ can be implemented as a scalable and viable alternative to handle secure communications in limited-capacity settings with recent additions that allow real-time tracking and visual diagnostics.
