An adversarial learning based image steganography with security improvement against neural network steganalysis

Steganographic schemes based on distortion functions are designed to embed messages while minimizing the desired distortion function. Conversely, steganalysis techniques like CNN-based methods attempt to identify the presence of a message in the image. This study presents a steganography method based on the application of adversarial learning theories to conventional distortion-based steganography systems. A function called saliency map, which gives each pixel a value based on its influence over the steganalysis classifier's output is employed to accomplish this. This number establishes the importance of each pixel in deceiving the steganalyzer. The image pixels are divided into two groups based on the saliency map to deceive the steganalyzer. Then, the two groups are combined by adjusting the embedding costs of the more significant group in line with the sign of the loss function. Additionally, a saliency map is used to manage the ratio of increasing or reducing embedding costs. The message is then embedded into the image using the traditional distortion-based steganography method S-UNIWARD while using new costs. Experimental results show that the suggested method can reduce the perturbation of produced stego images. Furthermore, the results reveal that 95.1% of generated images by the proposed method deceive the steganalyzer, indicating high performance compared to the baseline methods. In addition, we investigated the performance of the proposed method against an unaware and an aware steganalyzer.