AcousticWMD: Windows malware detection using machine learning models based on acoustic signal processing

Malware or malicious software is becoming increasingly complex, making traditional detection methods less effective. To address this issue, we propose a new method that transforms Windows executable files (Portable Executable or PE files) into sound signals. These signals are then analyzed using machine learning (ML) and deep learning (DL) techniques to detect malicious patterns. While this approach has been widely applied to Android application files (APK files), its use for detecting Windows PE malware remains under-investigated. In this study, we compare two types of audio features: MelFrequency Cepstral Coefficients (MFCC), which capture basic frequency patterns, and VGGish embeddings, which offer deeper semantic representations based on large-scale audio datasets. Our experimental results show that models trained on these features can achieve high detection accuracy, reaching up to $96.5 \%$ accuracy and an F1-score of $\mathbf{9 6. 8 \%}$. When VGGish embeddings are combined with a Multi-Layer Perceptron (MLP), an AUC score of $96.21 \%$ is achieved, highlighting the benefit of deep contextual features. Additionally, we evaluate three binary-toaudio conversion methods: Direct Binary Conversion, EntropyBased Conversion, and MIDI Conversion. Among them, Direct Binary Conversion provides the best trade-off between simplicity and performance. This research demonstrates that audio-based analysis powered by AI is a promising new direction for malware detection, with strong potential to uncover both known and sophisticated emerging threats.