A Graph-Based Multiple Instance Learning Framework for Steganographer Identification
Qianqian Zhang, Yi Zhang, Yuanyuan Ma, Ruiting Liu · 2024
Steganographer identification aims to identify malicious users who hide secret information in image. However, image-level annotation requires a significant amount of cost and time. To address these problems, we propose a novel algorithm for classifying users for steganographer identification based on user-level granularity without image-level annotation. Firstly, we point out that the image feature is one of the keys to the model, and present the conditions to be satisfied by the image feature. Based on this, we analyze and extract the features suitable for steganographer identification. Then, we take steganographer identification as a graph-based multi-instance learning problem, where an image is an instance, and a user is a bag containing multiple images. Next, we use the Relief-based weighting feature to describe the instance and construct a user graph by calculating the relationship between the instances to reduce noise interference. Finally, we train a graph kernel-based SVM classifier by measuring the similarity between each graph. The proposed method can improve the identification accuracy and robustness, and the results are interpretable. To evaluate the performance of the proposed method, we conduct a series of experiments on the BOSSBase-1.01 and BOWs2 dataset which commonly used in steganography and steganalysis. It can be observed from the experimental results that the proposed framework outperforms most of the compared methods and achieves better performance in identifying guilty users who use JPEG steganography.