Multi-task few-shot text steganalysis based on context-attentive prototypes

Abstract

The current mainstream text steganalysis methods are based on supervised training, requiring a sufficient number of training samples, and necessitating that the training and testing distributions meet the prerequisite of being independently and identically distributed (i.i.d.). However, in practical applications, the distribution of stego text may vary with factors such as text corpora, steganographic algorithms, and embedding rates. This issue, known as domain mismatch, will significantly impact the models’ detection performance. To address this issue, we propose a multi-task few-shot text steganalysis model based on Context-sensitive Prototypes, namely CP-Stega. CP-Stega first extracts generic and task-specific text features from the input sentences and constructs a prototype for each task by averaging the sentence representations of the support set. Next, CP-Stega measures the similarity between the query sample and each task by calculating the distance between the sentence representations and each prototype for any given query sample. To improve the model's adaptability to diverse tasks, we design a loss function consisting of multi-margin loss and KL divergence loss, which expands inter-task prototype distance and shortens intra-task distance, enabling the model to update its meta-parameters accordingly. Experiments reveal that our model can effectively adapt to diverse tasks under different meta-learning scenarios based on different embedding algorithms and embedding payloads, significantly improving text steganalysis performance compared to current foremost steganalysis models and other meta-learning algorithms. By calculating the distance distribution between the sentence representations and each prototype for any given query sample, CP-Stega effectively measures the similarity between the query sample and each task. To enhance the model's adaptability to diverse tasks, we design a loss function comprising multi-margin loss and KL divergence loss. The experiments demonstrate that CP-Stega can effectively perform various steganalysis detection tasks on three self-built meta datasets with different embedding algorithms and embedding capacities.