STEG-XAI: explainable steganalysis in images using neural networks

Multimedia content's development and technological evolution have enhanced and even facilitated the application of steganography as a means to introduce hidden messages for cybercrime-related purposes. Artificial intelligence models have been widely implemented as a way to detect the presence of these messages in image content. However, the possibility of applying explainability techniques in order to provide visual representations of the signatures of different steganography algorithms has not been studied yet. This work presents a novel steganalysys methodology, STEG-XAI, not only for detecting steganography in images but also for explaining the machine learning model's findings, and extracting the steganography algorithm's signature. A convolutional neural network with EfficientNet architecture is implemented, along with the explainability algorithms LIME and Grad-CAM. The model is trained with a dataset of images modified by UERD, a steganography method designed for JPEG images, and achieves a weighted AUC of 0.944, displaying a high level of discrimination between original and tampered images. Furthermore, the explanation methods enable visualizing both the image modifications identified by the neural network, and a signature of the UERD algorithm.