LIARS: Low-Information Area Recognition for Steganography

Steganography conceals information within digital media, making its detection a crucial challenge in modern digital forensics. This paper introduces a framework that integrates machine learning to automatically identify near uniform regions in images, to then embed data in them, and to validate the presence of such concealed content through a controlled forensic pipeline. The proposed workflow segments images into patches, classifies them as high-information or low-information using a convolutional neural network, and embeds a textual payload exclusively within low-entropy zones. A validation process based on difference-map and optical character recognition module are used to extract and confirm the hidden text. Experimental results demonstrate that the proposed work can reliably detect and reconstruct hidden payloads even when pixel intensity modifications are as subtle as ±5 RGB units, imperceptible to the human eye. The results highlight the potential of low-information modeling and adaptive AI segmentation as effective tools for explainable and forensic-oriented steganalysis.