Deep Learning for Steganographic Image Detection

Steganography is often difficult to detect, especially when applied to images, as alterations to the image are minimal and often imperceptible to the human eye. To address these obstacles, we trained EfficientNetV2-S, a Convolutional Neural Network (CNN) model optimized for efficiency, for steganography detection. Using a CNN allows for flexibility as it can train on subtle features created by steganographic methods and differentiate between them and clean images. We find that our fine-tuned model is able to differentiate between clean images and certain classes of steganographically altered images with an accuracy of up to 98 percent. Notably, our model is able to distinguish between four methods of steganographic alteration and clean images with a 66.29 percent accuracy, demonstrating a new capability in the field of steganalysis.