Is Audio Spoof Detection Robust to Laundering Attacks?

Authors: Hashim Ali, Surya Subramani, Shefali Sudhir, Raksha Varahamurthy, Hafiz Malik

Conference: ACM IH&MMSec 2024, Baiona, Spain

About

This study evaluates seven state-of-the-art audio spoof detection systems under various laundering attacks including reverberation, additive noise, recompression, resampling, and low-pass filtering. The proposed ASVSpoof Laundered Database extends the ASVSpoof 2019 LA eval partition by applying 33 distinct laundering conditions, resulting in 1388.22 hours of modified audio data.

Dataset Highlights

• 1388+ hours of processed audio
• 5 types of additive noise, 3 reverberation levels, 6 recompression bitrates, 4 resampling rates, and 1 low-pass filter
• Total Bonafide examples (across laundering attacks): 235,360
• Total Spoofed examples: 2,043,484

Results Snapshot

• Best Clean Condition: AASIST (0.83% EER)
• Worst under Noise: CQCC-GMM (avg EER 36.54–47.05%)
• Most robust to recompression: AASIST & RawGAT-ST (stable EER across bitrate changes)
• Most vulnerable: High RT60 reverberation & 44kHz resampling

Resources

• Explore the dataset on Hugging Face
• Access it via University of Michigan Deep Blue
• Read the full research paper (PDF)

How to cite:

@inproceedings{ali2024audio,
    title={Is Audio Spoof Detection Robust to Laundering Attacks?},
    author={Ali, Hashim and Subramani, Surya and Sudhir, Shefali and Varahamurthy, Raksha and Malik, Hafiz},
    booktitle={Proceedings of the 2024 ACM Workshop on Information Hiding and Multimedia Security},
    pages={283--288},
    year={2024}
}