Abstract: Existing image steganalysis methods based on deep learning assume that the embedding rates are known, whereas for most practical applications, these rates are unknown, leading to a sharp drop in model detection performance. This study combined transfer learning (TL) and multi-task learning (MTL) and proposed an image steganalysis method for a specific steganographic algorithm and unknown embedding rates. The proposed method used stego images with high embedding rates to pre-train the steganalysis model, constructed a steganalysis model based on MTL, and then transferred the parameter values of the pre-trained model as the initial values. The parameters were further fine-tuned on the training set, which consists of cover images and stego images with various embedding rates. A new objective function was designed by applying the weighting losses to the uncertainty method, dynamically adjusting the weight of each sub-task during the training process. The proposed method extracted the common features of images with various embedding rates more effectively, achieved better detection accuracy on images with unknown embedding rates, and demonstrated improved generalization ability.

Key words: steganalysis, transfer learning, multi-task learning, weighting losses with uncertainty