A Deep Learning Framework to identify real-world stego images

Image Steganalysis is the process of identifying images that have been processed using Steganography to hide information or messages. These images are called stego images. Identifying real-world stego images which vary in dimensions, scalability, lightning conditions with unknown steganography algorithms, payload, embedding capacity is a challenging task. This research proposes a deep learning framework that uses Convolutional Neural Network (CNN) to identify such real-world stego images. The framework consists of two pre-trained models namely InceptionNet V3 and EfficientNet B3. Models are trained using the ALASKA2 dataset and are evaluated using model accuracy. Additionally, evaluation metrics like Precision, Recall, and F1 scores are also calculated. Inception V3 achieves an accuracy of 73.33% while the performance of EfficientNet B3 is slightly better with an accuracy of 79.43%. Identification of real-world stego images will be of great benefit to the government security department to determine any illegal activity