Feature augmentation-based few-shot image steganalysis

Image steganalysis is a technique aimed at detecting and identifying secret information hidden within images. In recent years, supervised steganalysis methods based on deep learning have achieved good results. However, in practical applications, due to unknown steganographic algorithms and limited labeled samples, the performance of existing models significantly drops when detecting unknown samples. To address this issue, we propose a feature-augmented method for steganalysis with limited samples, aiming to improve the detection capability of models against unknown steganographic algorithms using a small number of labeled samples. The proposed architecture consists of two main components: stego sample augmentation and model hierarchical training. First, in the stego sample augmentation phase, high-frequency noise features of images generated by unknown steganographic algorithms are captured using spatial-rich models and 2D Gabor filter kernels. These features are then processed by a normalized feature saliency metric to select the most representative high-frequency stego feature maps. These feature maps are used to adjust a variational autoencoder model to better learn and construct prototypes of steganographic features. By sampling in the latent space of these prototypes and using a decoder, threshold segmentation, and an embedding simulator, a large number of pseudo-stego noise images are generated. These are then fused with denoised images to synthesize pseudo-stego samples. In the hierarchical model training phase, these synthesized pseudo-stego samples, along with real stego samples, are used to fine-tune a pretrained model. This involves coarse and fine adjustments to progressively optimize model performance, thereby enhancing the detection accuracy for unknown steganographic algorithms with limited labeled data. The experimental results show that this method demonstrates significant improvements in detecting unknown steganographic algorithms across various baseline steganalysis models while also performing well in detecting adversarial steganographic samples. Furthermore, ablation experiments validate the rationality of the proposed method and the effectiveness of its key components.