Unknown Performance of the APVD Reverse-Steganalysis Model on Color Images

Determine the detection accuracy, precision, recall, F1-score, Bit Error Rate, and payload recovery rate of the convolutional neural network with attention and dual output heads proposed for reverse steganalysis of Adaptive Pixel Value Differencing (APVD) stego-images when applied to color images, across varying payload sizes and APVD configurations.

Background

The paper introduces a dual-head convolutional neural network (CNN) with attention mechanisms to perform both detection and payload reconstruction for images steganographed using Adaptive Pixel Value Differencing (APVD). All experiments were conducted on grayscale images sourced from BOSSbase and UCID, yielding high detection accuracy and promising payload recovery at lower embedding rates.

Despite these results, the authors explicitly note that the model has not been evaluated on color images, leaving its performance in RGB or other color spaces unresolved. This gap is significant because APVD-based embedding and corresponding artifacts may manifest differently across color channels, which could affect both detection and reconstruction outcomes in real-world forensic scenarios.