Blind audio steganalysis of stego-malware attacks using harmonic-aware deep attention-based learning architecture

Stego-malware in compressed audio poses a covert threat by hiding commands in imperceptible regions, bypassing traditional detection methods. Hence, a novel, Harmonic-Aware Bidirectional Attention Network is proposed to overcome these issues and classify normal audio from stego audio. Initially, existing feature extractors suffer from quefrency aliasing and the loss of fine-grained phase and harmonic details, making them inadequate for identifying subtle steganographic cues in polyphonic environments. Thus, Harmonic-Chroma-Q Transform Feature Extractor (HCQT-FE) is introduced, which uses Harmonic-Percussive Source Separation (HPSS) for isolating tonal elements, Mel-Chroma-Spectral Contrast for enriching spectral and harmonic representations, and Constant-Q Transform (CQT) to preserve log-scaled frequency details aligned with audio structure, thereby capturing multi-dimensional audio representations. Moreover. traditional deep learning models struggle with polyphonic audio due to their inability to capture long-range dependencies amid overlapping harmonic structures, asynchronous onsets, and fluctuating tonal centers. Therefore, a Bidirectional Gated Attention Network (Bi-GAttNet) is presented, which uses Bidirectional LSTMs to model temporal context in both directions and a Gated Attention mechanism to focus on salient harmonic regions while suppressing noise dynamically enhances the ability to detect and classify covert steganographic patterns embedded within audio, achieving improved accuracy, recall, and robustness in stego-malware detection tasks.