Image steganalysis using reinforcement learning-based active learning and scope loss function

Image steganalysis, which involves detecting concealed data within image files, is crucial for enhancing digital security. Traditional methods rely on deep convolutional neural networks (CNNs) and require large, pre-labeled image datasets, which are time-consuming and costly to compile. Addressing this challenge, we introduce an innovative active learning approach that improves model performance with fewer labeled samples. Conventional active learning techniques often use static selection methods that limit adaptability and fail to optimize for dynamic environments. To overcome this, our approach incorporates deep reinforcement learning (DRL) enhanced by a novel scope loss function, balancing exploiting known data with exploring new data opportunities. The proposed model employs multiple CNNs to extract feature vectors from various layers of an image's structure, which are then processed by fully connected (FC) layers to categorize each image. Initially, the classifier is trained with a modest amount of labeled data. As the algorithm progresses, the DRL determines which unlabeled images should be labeled. Annotated images are added to the labeled data pool, and the classifier is periodically retrained to improve accuracy. Additionally, the model uses an improved differential evolution (DE) algorithm for complex hyperparameter management, using a new mutation mechanism with k-means clustering to identify significant clusters. Testing on the BossBase 1.01 and BOWS-2 datasets shows the method's ability to distinguish between unaltered and steganographic images, achieving average F-measures of 92.973% and 90.778%, respectively. This research advances digital security by enhancing image steganalysis techniques, significantly improving detection accuracy with limited labeled data.