About

This repo contains the ENF-WHU audio recording dataset collected around Wuhan University campus and the MATLAB programs for electronic network frequency (ENF) detection, enhancement, and robust estimation, in ENF-based audio forensic applications.

Note about the Ground-Truth (April 2023)

The ground-truth matched location (the lag that corresponding to the true timestamp) within the one day reference can be obtained by matching the noise-free ref files with the corresponding one day ref. For example, we can match "003_ref.wav" in "H1_ref" folder within "003-004_ref.wav" in "H1_ref_one_day" folder, and the matched lag index is the "ground truth" timestamp for recording "003.wav" in "H1" folder, meaning that "003.wav" should be matched at the same or a very close lag index in "003-004_ref.wav". Both MSE and CC can be used for the matching criterion as long as the recording and ref are matched using the same criterion.

ENF-WHU Dataset

• Recording location: classroom, campus path, meeting room, graduate student office, dormitory, library.
• Environment diversity: day/night, rainy/suny, interior/exterior.
• Recording device: popular smartphone and voice recorder.
• Duration: 5~20 minutes
• Format: PCM WAVE
• Quantization depth: 16-bit
• Channel: mono
• Sampling frequencuy: 8000 Hz (400 Hz for reference data)
• Category:
  H1: "001~130.wav" 130 real-world recordings with captured (noisy) ENF.
  H1_ref: "001_ref~130_ref.wav" the corresponding 130 reference ENF (noise-free, same duration) obtained from power main.
  H1_ref_one_day: the corresponding one-day (24 hours) reference ENF for the 130 recordings. "003-004_ref.wav" means "003.wav" and "004.wav" in H1 are recorded within the same day.
  H0: "O01~O10.wav" 10 real-world recordings without captured ENF. "01~40.wav" 40 segments under H0 obtained by random cropping the 10 recordings.

MATLAB Programs

ENF Detection

• Clairvoyant detectors: NP detectors assuming perfect knowledge of ENF.
  1. GMF: a standard NP detector.
  2. MF-like approximation: avoid the requirement of unknown noise covariance matrix.
  3. Asymptotic approximation: trade-off between computational complexity and detection performance.
• GLRT detectors: ENF assumed unknown and deterministic.
  1. LS-LRT: MF-like with unknown parameters replaced by the MLEs.
  2. naive-LRT: MF-like with the unknown IFs replaced by nominal value.
• TF domain detector: ENF assumed unknown and random.
  • Test statistic is the sample variance of the strongest time-frequency line (e.g., STFT + peak)
  • Exploiting slow-varying nature of ENF, thus test statistic is large under H0 and small under H1.

ENF Enhancement and Estimation

It contains our proposed ENF enhancement and estimation methods including

• Single-tone model based ENF enhancement method [3],
• Multi-tone harmonic model based enhancement and harmonic selection for robust ENF estimation [2],
  

in comparison with the following existing works

• Adaptive multi-trace carving (AMTC) method by Q. Zhu et al. 2021 IEEE TIFS,
• Multi-tone maximum likelihood estimator (MLE) by Bykhovsky and Cohen 2013 IEEE TIFS,
• Multi-tone weighted maximum likelihood estimator (WMLE) by Hajj-Ahmad et al. 2013 IEEE SPL,
  

evaluated using both synthetic data and the real-world recordings from the ENF-WHU dataset.

Citation Information

• ENF Detection:

[1] G. Hua, H. Liao, Q. Wang, H. Zhang, and D. Ye, "Detection of electric network frequency in audio recordings – From theory to practical detectors," IEEE Trans. Inf. Forensics Security, vol. 16, pp. 236–248, 2021. link

• ENF Enhancement:

[2] G. Hua, H. Liao, H. Zhang, D. Ye, and J. Ma, "Robust ENF estimation based on harmonic enhancement and maximum weight clique," IEEE Trans. Inf. Forensics Security, DOI: 10.1109/TIFS.2021.3099697, 2021. link
[3] G. Hua and H. Zhang, "ENF signal enhancement in audio recordings," IEEE Trans. Inf. Forensics Security, vol. 15, pp. 1868-1878, 2020. link

• Related Works:

[4] G. Hua, "Error analysis of forensic ENF matching," in Proc. 2018 IEEE International Workshop on Information Forensics and Security (WIFS), pp. 1-7, Hong Kong, Dec. 2018. link
[5] G. Hua, G. Bi, and V. L. L. Thing, "On practical issues of electric network frequency based audio forensics," IEEE Access, vol. 5, pp. 20640-20651, Oct. 2017. link
[6] G. Hua, Y. Zhang, J. Goh, and V. L. L. Thing, "Audio authentication by exploring the absolute error map of the ENF signals," IEEE Trans. Inf. Forensics Security, vol. 11, no. 5, pp. 1003-1016, May 2016. link
[7] G. Hua, J. Goh, and V. L. L. Thing, “A dynamic matching algorithm for audio timestamp identification using the ENF criterion,” IEEE Trans. Inf. Forensics Security, vol. 9, no. 7, pp. 1045-1055, Jul. 2014. link