Steganalysis of Adaptive Multi-Rate Speech with Unknown Embedding Rates Using Multi-Scale Transformer and Multi-Task Learning Mechanism

Abstract

As adaptive multi-rate (AMR) speech applications become increasingly widespread, AMR-based steganography presents growing security risks. Conventional steganalysis methods often assume known embedding rates, limiting their practicality in real-world scenarios where embedding rates are unknown. To overcome this limitation, we introduce a novel framework that integrates a multi-scale transformer architecture with multi-task learning for joint classification and regression. The classification task effectively distinguishes between cover and stego samples, while the regression task enhances feature representation by predicting continuous embedding values, providing deeper insights into embedding behaviors. This joint optimization strategy improves model adaptability to diverse embedding conditions and captures the underlying relationships between discrete embedding classes and their continuous distributions. The experimental results demonstrate that our approach achieves higher accuracy and robustness than existing steganalysis methods across varying embedding rates.