Stego Guard: A Residual Cnn-Based Framework for Image Steganalysis and Browser-Integrated Detection

Abstract

The increasing use of digital platforms has made images one of the most frequently exchanged forms of online content. Alongside legitimate communication, images can also be employed as carriers for concealed information through steganographic techniques. Steganography attempts to hide the existence of a secret message by embedding it within a supposedly ordinary digital medium. This characteristic makes the identification of hidden information an important problem in digital image security. The proposed StegoGuard framework addresses this problem using a Residual Convolutional Neural Network (CNN) for image steganalysis. The system divides input images into 128 × 128 pixel patches and evaluates these regions individually using the trained Residual CNN. The resulting patch-level predictions are used to obtain an image-level stego score, which is compared against a threshold of 0.66 to classify the image as CLEAN or STEGO. The detection model is integrated with a browser extension through Native Messaging, allowing images encountered during web browsing to be analysed through a local detection application. The experimental evaluation achieved an accuracy of approximately 96% on the dataset used in this study. The results demonstrate the potential of combining Residual CNN-based analysis, patch-level detection, and browser integration for practical image steganalysis. The presence of false-positive classifications also indicates the need for more diverse datasets and broader evaluation.