bellingcat/open-questions
1
0
Fork 0
mirror of https://github.com/bellingcat/open-questions.git synced 2026-06-07 19:08:30 +03:00
Code Issues Packages Projects Releases Wiki Activity

Added papers that started ENF signal extraction and verification from audio signals.

Browse Source
This commit is contained in:
Radek Ludačka
2021-07-24 20:08:57 +02:00
parent 7a2ac387d1
commit 6b42f2e3a7
2 changed files with 40 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

39
electrical-network-frequency-analysis/papers.md
View File

@@ -1,5 +1,44 @@
# Papers
###### Digital Audio Recording Analysis The Electric Network Frequency Criterion
Probably first paper about ENF extraction from audio signals.
Using Frequency Domain Analysis.
Suggested approach:
* Using DCLive Forensics software.
* Down sample the evidence to 120 Hz. (with "Change Sample Rate / Resolution" function)
* Band-pass filter the audio with cut frequencies around 49 and 51 Hz (with "Band Pass Filter")
* Compute the spectrogram with a 4096 points FFT.
* Make a vertical zoom around 50 Hz.
* Compare with the ENF database. (Probably manually - eye balling spectrogram).
https://www.tracertek.com/media/pdf/an4.pdf
###### Applications of ENF criterion in forensic audio, video, computer and telecommunication analysis
Probably first paper about ENF extraction from audio signals describing 3 possible methods.
* Time/frequency domain spectrograms method consists on computing spectrograms
* Visually compare questioned versus database ENF.
* Useful especially for questioned versus database ENF date and time verification.
* The fastest one (if we know exect recording time), is easy to be implemented.
* Reveals the ENF components number.
* For longer than 10–15 min recordings.
* Indicates a non-authentic or non-duplicate evidence recording.
* Must be used before the other two in order to find out the ENF components number.
* Frequency domain
* Method means to compute FFT over short time windows.
* Extract the maximum magnitude value around 50/60 Hz and compare questioned samples versus database ENF.
* Can be applied to one ENF component recordings for questioned versus database ENF verification or for database searching in order to identify the questioned ENF recording date and time.
* Can also be used on more than one ENF component recordings in order to separate the ENF components and detect their chronological enter the evidence.
* Time domain analysis
* Consists on zero-crosses measure.
* One ENF component recordings only.
* For bigger than 44 kHz sampling frequencies and by implementing zero-crosses interpolation can become sampling frequency independent.
* Can offer more useful detail than a frequency domain analysis.
* Applying a band-pass filter with 49–51 Hz cut frequencies, without down sampling the signal, one can separate the ENF waveform from the rest of the recording.
* Possible to analyse the original quantification levels of the signal.
* Calculation degree of correlation between questioned ENF component and reference database ENF for search/verification purpose is used for the last two methods.
https://www.sciencedirect.com/science/article/abs/pii/S0379073806004312
###### Power Grid Estimation Using Electric Network Frequency Signals
Using XGBoost classifier with 95.21% and 99.07% precision when ENF signals have 480 and 1920 data points
https://www.hindawi.com/journals/scn/2019/1982168/

2
electrical-network-frequency-analysis/useable-source-codes.md
View File

@@ -2,7 +2,7 @@
### PyEnf
* Python based, ENF extraction from audio recordings.
* Using harmonic frequencies and frequency spectorgram.
* Using harmonic frequencies and frequency spectrogram.
https://github.com/deerajnagothu/pyenf_extraction