Enhancing Secret Data Detection Using Convolutional Neural Networks With Fuzzy Edge Detection

Progress in Deep Learning (DL) has introduced alternative methods for tackling complex challenges, such as the steganalysis of spatial domain images, where Convolutional Neural Networks (CNNs) are employed. In recent years, various CNN architectures have emerged, enhancing the precision of detecting steganographic images. Nevertheless, current CNNs encounter challenges related to the inadequate quality and quantity of available datasets, high imperceptibility of low payload capacities, and suboptimal feature learning processes. This paper proposes an enhanced secret data detection approach with a CNN architecture that includes convolutional, depth-wise, separable, pooling, and spatial dropout layers. An improved fuzzy Prewitt approach is employed for pre-processing the images prior to being fed into CNN to address the issues of low payload capacity detection and dataset quality and quantity in learnability of the image features. Experimental results, which achieved an overall accuracy and F1-score of 99.6 and 99.3 per cent, respectively, to detect a steganographic payload of 0.5 bpp hidden with Wavelet Obtained Weights (WOW), show a significant outperformance over the state-of-the-art methods.