Design of a integrated steganalysis model for adversarial steganography

Currently, although deep learning based steganalysis methods have shown significant advantages in detection performance compared to traditional methods, they are highly susceptible to attacks from adversarial steganalysis methods. How to synergistically leverage the advantages of two methods in steganalysis tasks has become a key issue that urgently needs to be addressed. Based on this, a fused steganalysis framework is proposed. The method uses the traditional steganalysis method based on SRM (Spatial Rich Model) manual features and the deep learning steganalysis method Ye Net as base learners, and integrates their discriminative outputs through ensemble learning; At the same time, a deep classifier based on adversarial transfer network is constructed, which relies on the adversarial game process of feature extractor and domain discriminator to extract domain invariant features that can be shared between non adversarial and adversarial domains, achieving effective model training in scenarios where the true value label of the adversarial domain is unknown. In addition, the model is based on MLP (Multi Layer Perceptron) to construct a deviation sample recognition module, effectively suppressing the negative transfer phenomenon that occurs during the training process, stabilizing the domain distribution alignment process, and further improving the model's cross domain generalization ability in adversarial disturbance environments. The experimental results showed that under different embedding rates and intensities of adversarial steganalysis attacks, the proposed fusion steganalysis model had an average decrease of 15.95% and 6.06% in error rate [M2] (Probability of Error, Pe) compared to the traditional steganalysis models SPAM (Subtractive Pixel Adjacency Matrix steganalysis) and SRM [M1], and an average decrease of 10.93% to 19.50% in error rate compared to deep learning steganalysis models (Ye Net, SRNet, LWENet). Compared to the robustness enhancement method KDNFT (K-times Dropout Neighboring Feature Transformer) [M3] for adversarial steganalysis, the error rate decreased by an average of 5.90%. Achieve the steganalysis performance of the current SOTA in adversarial steganography scenarios. Conclusion: The fusion based steganalysis framework proposed in this article can effectively reduce the comprehensive error rate of detecting adversarial sample steganographic images, providing a new feasible path for achieving more accurate and highly reliable steganalysis models.