[Paper Review] Spec-ResNet: A General Audio Steganalysis scheme based on Deep Residual Network of Spectrogram

Yanzhen Ren, Dengkai Liu+3 more|arXiv (Cornell University)|Jan 21, 2019

Digital Media Forensic Detection37 references17 citations

TL;DR

This paper proposes Spec-ResNet, a general audio steganalysis framework that uses spectrograms as input and a deep residual neural network to detect steganography in AAC and MP3 audio across multiple embedding domains. By capturing universal steganalysis features through multi-scale spectrogram windows and leveraging residual learning, the method achieves state-of-the-art detection accuracy, outperforming existing hand-crafted and CNN-based approaches on AAC and MP3Stego schemes.

ABSTRACT

The widespread application of audio and video communication technology make the compressed audio data flowing over the Internet, and make it become an important carrier for covert communication. There are many steganographic schemes emerged in the mainstream audio compression data, such as AAC and MP3, followed by many steganalysis schemes. However, these steganalysis schemes are only effective in the specific embedded domain. In this paper, a general steganalysis scheme Spec-ResNet (Deep Residual Network of Spectrogram) is proposed to detect the steganography schemes of different embedding domain for AAC and MP3. The basic idea is that the steganographic modification of different embedding domain will all introduce the change of the decoded audio signal. In this paper, the spectrogram, which is the visual representation of the spectrum of frequencies of audio signal, is adopted as the input of the feature network to extract the universal features introduced by steganography schemes; Deep Neural Network Spec-ResNet is well-designed to represent the steganalysis feature; and the features extracted from different spectrogram windows are combined to fully capture the steganalysis features. The experiment results show that the proposed scheme has good detection accuracy and generality. The proposed scheme has better detection accuracy for three different AAC steganographic schemes and MP3Stego than the state-of-arts steganalysis schemes which are based on traditional hand-crafted or CNN-based feature. To the best of our knowledge, the audio steganalysis scheme based on the spectrogram and deep residual network is first proposed in this paper. The method proposed in this paper can be extended to the audio steganalysis of other codec or audio forensics.

Motivation & Objective

• To develop a generalizable steganalysis framework capable of detecting steganography across diverse embedding domains in compressed audio formats like AAC and MP3.
• To overcome the limitations of existing steganalysis methods that are effective only within specific embedding domains.
• To explore the use of spectrograms as a universal representation for capturing subtle steganographic modifications in audio signals.
• To leverage deep residual networks to effectively extract weak steganalysis features from spectrogram representations.
• To improve detection accuracy through multi-scale feature fusion from different spectrogram window sizes.

Proposed method

• The method uses spectrograms—visual representations of audio frequency content—as input to the deep neural network, capturing spectral-temporal relationships in compressed audio.
• A deep residual network (ResNet) is employed to learn hierarchical steganalysis features, mitigating vanishing gradient issues common in deep architectures.
• The model processes spectrograms from multiple window sizes (N=512, 1024, 2048) to extract multi-scale features, which are then fused for improved detection.
• Training is performed on spectrograms from 1500 audio samples in CDB_MP3 and SDB_MP3 datasets, with separate training and test sets to avoid data leakage.
• The model is trained end-to-end using cross-entropy loss and optimized via stochastic gradient descent with learning rate decay.
• A mixture classification model is constructed by combining predictions from multiple window sizes to enhance robustness and accuracy.

Experimental results

Research questions

• RQ1 Can spectrograms serve as a universal feature representation for detecting steganography across different audio compression domains?
• RQ2 Can a deep residual network effectively learn subtle steganographic distortions in compressed audio signals?
• RQ3 Does multi-scale feature fusion from spectrograms improve steganalysis detection accuracy compared to single-scale analysis?
• RQ4 How does the proposed Spec-ResNet method compare to state-of-the-art steganalysis schemes in terms of accuracy and generalization across AAC and MP3 steganography schemes?
• RQ5 Is the proposed method generalizable to other audio codecs or forensic applications beyond steganalysis?

Key findings

• Spec-ResNet achieved a detection accuracy of 0.9993 on MP3Stego using a mixture model with N=1024 window size, significantly outperforming Wang [25] (0.9250) and Luo [32] (0.9786).
• For AAC steganography, Spec-ResNet achieved an average detection accuracy of 0.8829 across three schemes (LSB-EE, MIN, SIGN), surpassing prior state-of-the-art methods.
• The model with N=1024 window size achieved 100% true positive rate (TPR) and 99.87% true negative rate (TNR) on MP3Stego, indicating near-perfect detection performance.
• The mixture model combining multiple window sizes achieved the highest accuracy (0.9993), demonstrating the effectiveness of multi-scale feature fusion.
• Spec-ResNet achieved 0.9531 detection accuracy on the LSB-EE scheme in AAC, outperforming the next best method (0.8819) by over 7 percentage points.
• To the best of the authors' knowledge, this is the first published work combining spectrograms and deep residual networks for audio steganalysis, establishing a new benchmark in the field.