Frame-Wise Steganalysis of AMR Encoded Speech Streams Using Automatic Codeword Feature Interaction Mechanism

Congcong Sun, Hui Tian, Peng Tian, Haizhou Li, Ching-Chun Chang, Chin-Chen ChangIEEE Transactions on Audio, Speech and Language Processing 2025 · OpenReview

Abstract

Frame-wise steganalysis is crucial in AMR-encoded speech streams. By analyzing each frame individually, it becomes possible to precisely identify which frame has been inserted into confidential Information and easily alter the codewords in the frame to prevent the confidential information from being extracted correctly. However, no common interpretable theory has been proposed for frame-wise detection in AMR-encoded speech streams. Moreover, the existing frame-wise steganalysis methods cannot fully exploit the correlation between the codewords in each frame. Therefore, a frame-wise steganalysis method based on the automatic codeword feature interaction mechanism is proposed in this paper. The automatic codeword feature interaction mechanism can be achieved through multiple iterations. During these iterations, high-order codeword interaction features are obtained. Each iteration consists of two steps. The first step employs an element-wise codeword feature interaction mechanism to capture fine-grained, localized patterns between corresponding elements. The second step utilizes a vector-wise codeword feature interaction mechanism to capture global patterns across entire codeword vectors. To enhance the learning process and facilitate better information flow, a residual connection is added to the automatic codeword feature interaction mechanism. On numerous large steganography datasets, the experimental results reveal that our method has a significant improvement for the frame-wise steganalysis task, compared to the existing frame-wise steganalysis methods.
BibTeX
@article{sun2025frame,
  author = {Congcong Sun and Hui Tian and Peng Tian and Haizhou Li and Ching-Chun Chang and Chin-Chen Chang},
  title = {{Frame-Wise Steganalysis of AMR Encoded Speech Streams Using Automatic Codeword Feature Interaction Mechanism}},
  journal = {IEEE Transactions on Audio, Speech and Language Processing},
  year = {2025},
  url = {https://openreview.net/forum?id=Eg8lOMACaJ},
}

Loading related content...

Figure