Fuzzy-Clustering-Based Domain Adaptation for Speech Steganalysis in Dynamic Scenarios

Rong Fu, Lili Tang, Hui Tian, Chin-Chen Chang

Abstract: Existing speech steganalysis suffers from generalization in dynamic scenarios such as unknown or random Embedding Strengths (ES). The central challenge is domain mismatch, a problem that remains largely unexplored in speech steganalysis. To fill the gap, this letter proposes a novel multi-source domain adaptation method called Fuzzy-Clustering-Based Domain Adaptation (FCDA). First, to enable effective clustering of samples with similar actual ES, FCDA employs Fuzzy C-Means (FCM) clustering, allowing soft estimation across multiple Embedding Rate (ER) levels. Second, to enhance the steganographic classification performance, we construct the backbone network integrating a dual-primary classifier with an auxiliary ER hierarchy classifier, which leverages the strength-sensitive representation. Third, to better realize the training of multi-objective tasks and enhance feature discriminability, the domain-consistent optimization loss is introduced. Experimental results show that FCDA outperforms state-of-the-art steganalysis methods and multi-source domain adaptation methods.