INSTANTANEOUS FREQUENCY ESTIMATION AND LOCALIZATION FOR ENF SIGNALS

Adi Hajj-Ahmad, Ravi Garg, and Min Wu

{adiha, ravig, minwu}@umd.edu

Department of Electrical & Computer Engineering
Institute for Advanced Computer Studies
University of Maryland, College Park, MD, USA.

ABSTRACT timestamp authentication and forgery detection in multimedia

recordings.
Forensic analysis based on Electric Network Frequency
(ENF) fluctuations is an emerging technology for authenticat- The ENF signal is embedded in such multimedia record-
ing multimedia recordings. This class of techniques requires ings as audio and video at the time when the recording is
extracting frequency fluctuations from multimedia recordings conducted. Audio recordings made near power source can
and comparing them with the ground truth frequencies, ob- capture the ENF variations from the electromagnetic interfer-
tained from the power mains, at the corresponding time. Most ence from power lines [1]. Similarly, indoor video recordings
current guidelines for frequency estimation from the ENF made under fluorescent or incandescent lightings can cap-
signal use non-parametric approaches. Such approaches have ture the ENF signal due to near invisible flickering of light-
limited temporal-frequency resolution due to the tradeoffs ings [2]. Frequency estimation techniques can then be applied
of the time-frequency resolutions as well as computational to extract the ENF signals from media files for forensic stud-
power. To facilitate robust high-resolution matching, it is ies. One such forensic application is the determination of the
important to estimate instantaneous frequency using as few time-of-recording of a given file by comparing the ENF signal
samples as possible. The use of subspace-based methods for extracted from the media file with the ENF signal obtained di-
high resolution frequency estimation is fairly new for ENF rectly from the power mains. In another application, the ENF
analysis. In this paper, a systematic study of several high signals from audio and video tracks of a given video can be
resolution low-complexity frequency estimation algorithms compared to determine if the audio and visual data were cap-
is conducted, focusing on estimating the frequencies in short tured at the same time [2].
time-frames. After establishing the performance of several Accurately and reliably estimating ENF signals is of ut-
frequency estimation algorithms, a study towards using the most importance in its application for forensic purposes. This
ENF signal for estimating the location-of-recording is carried is because the validity of results depends on the accuracy of
out. Experiments conducted on ENF data collected in several the frequency estimates obtained using a relatively short seg-
cities indicate the presence of location-specific signatures that ment of media and sensor recordings. ENF signals in different
can be exploited for future forensic applications. recordings may be of varying quality. For example, ENF sig-
nals embedded in audio recordings are of low signal-to-noise
Index Terms— Electric Network Frequency (ENF), ratio as compared with the signal recorded directly from the
Robust Instantaneous Frequency Estimation, Localization, power mains. Moreover, the temporal resolution in timestamp
Location-stamp estimation in media recordings using ENF signals is limited
by the duration of frames used for instantaneous frequency es-
1. INTRODUCTION timation. Using a short frame may improve the resolution at a
cost of decreased reliability in obtaining frequency estimates,
Analysis of the Electric Network Frequency (ENF) signal is a if an appropriate frequency estimation is not used. Similarly,
recently proposed tool for multimedia forensics [1] [2]. ENF limited temporal resolution may also affect the capabilities of
is the supply frequency in power distribution networks and its a forensic examiner analyzing a possible forgery on ENF sig-
nominal value is 50 or 60 Hz depending on the geographic nals for different counter forensic scenarios [3].
location. An important property of the ENF signal is that its In the first part of this paper, we examine several tech-
value fluctuates around the nominal value; for example, the niques for estimating the instantaneous ENF frequencies in
range of fluctuations is roughly on the order of 50-100 mHz a raw measurement signal under different noise conditions.
in the United States. These fluctuations are due to variations We also examine the effect of varying the frame size on the
in the load on the power grid and generally can be considered reliability of ENF signal estimation. The number of frames
as random. This property of the ENF signal has enabled its can become quite large for long media files, so we focus our
emerging use in multimedia forensic analysis, particularly for study on examining methods without overly high computa-
tional complexity. We study the performance of different fre- section are based on the spectrogram and the time recursive
quency estimation methods on synthetic and power data. iterative adaptive approach (TR-IAA).
After experimentally evaluating the performance of dif- The spectrogram is a widely used tool for visualizing the
ferent frequency estimation methods, we explore a novel ap- Short-Time Fourier Transform (STFT). STFT is usually used
plication of ENF signal analysis in pinpointing the location for signals with a time-varying spectrum, such as the ENF
in which a media or sensor recording was done. Most of the signal. After the signal is segmented into overlapping frames,
existing work has established that the ENF signal across an each frame undergoes Fourier analysis to determine the fre-
interconnected power grid is similar at a given time period. quencies present. The spectrogram is then defined as the
This is due to the load balancing mechanism employed in the squared magnitude of the STFT. It is usually displayed as a
power grids [4]. Variations may still be present in the fre- two-dimensional intensity plot, with the two axes being time
quency fluctuations at different locations due to local changes and frequency.
in the load and the finite propagation speed of the effects of The ENF signal in a frame can be considered to be a si-
this load change to other parts of the grid [5]. In this pa- nusoid of a frequency close to the nominal value, embedded
per, we study these effects by conducting experiments on ENF in noise. So, the power spectral density (PSD) of it, esti-
data collected across several locations within the Eastern grid mated by the spectrogram method, should ideally have a peak
of the United States. This is the first study exploiting the ENF at the frequency of the sinusoidal signal. We consider two
signal in location-stamping. It will be shown in our prelim- spectrogram-based approaches: (a) the maximum energy ap-
inary analysis that there exists differences among recordings proach with quadratic interpolation, and (b) the weighted en-
taken in different locations within the same interconnected ergy approach.
power grid. Such a property of the ENF signal may facili-
tate its use in a new range of forensic applications to pinpoint Maximum energy with quadratic interpolation: In this ap-
the location where a media file was recorded. The location proach, we begin with finding the frequency that has the max-
specific traces in the ENF signal may be useful in an efficient imum spectral power component. Directly choosing this fre-
monitoring of future smart-grids by assisting in fault localiza- quency as the frame’s frequency estimate leads to much loss
tion [6]. in accuracy, because the spectrogram is computed for a finite
The remaining of the paper is organized as follows. In number of discrete frequencies and the actual frequency with
Section 2, we study different frequency estimation techniques the maximum energy may not be among them. To improve the
for ENF signal analysis. In Section 3, we experimentally accuracy of the estimate, the number of Fast Fourier Trans-
evaluate the performance of our examined methods for ENF form (FFT) points can be increased at a higher computational
signal estimation under different signal quality conditions and cost. An alternative approach is to use quadratic interpolation
different frame sizes. In Section 4, we study the characteris- for fitting the PSD points corresponding to the range about
tics of ENF signals obtained from different locations and pro- the discrete frequency with the maximum energy [7]. This
pose a method to extract location specific signatures. Section interpolation is outlined below based on the approach in [8].
5 concludes the paper. As the computed PSD of each frame is a function of the
discrete frequencies, we can think of it as a function of fre-
quency bin numbers, or simply indices. Denoting the index
2. ENF EXTRACTION METHODS of the frequency with the maximum energy as 𝑘𝑚𝑎𝑥 , we can
define a coordinate system centered at (𝑘𝑚𝑎𝑥 , 0). We take the
For ENF extraction, a source signal containing the ENF sig- log magnitude value of the PSD as 𝑦(𝑘). Using three points
nal is divided into overlapping frames of the same size. A fre- of the PSD, namely at 𝑘𝑚𝑎𝑥 − 1, 𝑘𝑚𝑎𝑥 and 𝑘𝑚𝑎𝑥 + 1, we can
quency estimation algorithm is used on each frame to estimate carry out quadratic interpolation on the parabola of the form:
the instantaneous frequency present in the frame. The result 𝑦(𝑘) = 𝑎(𝑘 − 𝑝)2 + 𝑏. Solving for parabola peak 𝑝, we get
would be a “frequency vector”, containing the frequency esti- the following expression.
mate for each subsequent frame. Retrieving frequency vectors
from power signals and audio/video signals is an important 1 𝑦(𝑘𝑚𝑎𝑥 − 1) − 𝑦(𝑘𝑚𝑎𝑥 + 1)
𝑝= , (1)
first step in the ENF analysis. In this section, we review and 2 𝑦(𝑘𝑚𝑎𝑥 − 1) − 2𝑦(𝑘𝑚𝑎𝑥 ) + 𝑦(𝑘𝑚𝑎𝑥 + 1)
discuss a few estimation techniques. Broadly speaking, they
fall into one of two categories: the non-parametric approaches where 𝑦(𝑘) is defined as
and the parametric subspace-based approaches. 𝑦(𝑘) = 20 log10 ∣𝑃 𝑆𝐷(𝑘)∣. (2)

2.1. Non-Parametric Approaches The estimate of the peak location in bins corresponding
to the true frequency is 𝑘𝑡 = 𝑘𝑚𝑎𝑥 + 𝑝. The frequency es-
Non-parametric approaches do not assume any explicit model timate corresponding to 𝑘𝑡 is 𝑘𝑁
𝑡 𝑓𝑠
, where 𝑓𝑠 is the sampling
for the data. Most of such approaches are based on the Fourier frequency and 𝑁 the number of FFT points used in computing
analysis of a signal. The main methods considered in this the spectrogram.
 A drawback of this approach is that it is susceptible to white noise. As ENF signals consist of only one real sinu-
outliers. If the maximum energy happens to be far away from soid, the value of 𝑃 for ENF signals is 2. The methods we
the nominal frequency due to additive noise or interference study are the MUltiple SIgnal Classification (MUSIC) and Es-
from content, the subsequent estimation would be erroneous. timation of Signal Parameters via Rotational Invariance Tech-
niques (ESPRIT). Next, we briefly describe these two meth-
Weighted energy: The weighted energy approach makes ods [10].
use of the additional information that we have, namely that
the frequency should be close to the nominal frequency. As MUSIC: The MUSIC algorithm is a subspace-based ap-
a result, the frequency estimates are robust to outliers in fre- proach to frequency estimation that relies on eigen-decomposition
quency [2]. and the properties between the signal and noise subspaces for
The weighted energy approach finds the frequency esti- sinusoidal signals with additive white noise. The algorithm
mate 𝐹 (𝑛), for the 𝑛𝑡ℎ frame, by weighing the frequency bins first computes an 𝑀 × 𝑀 correlation matrix, where 𝑀 is
around the nominal values. The expression for frequency es- chosen to be larger than 𝑃 , the number of anticipated com-
timates is then given by the following equation [2]: plex exponentials. More specifically, for an 𝑁 -point observed
∑𝐿2 signal 𝑥[𝑛] where 𝑁 >> 𝑀 , we generate an 𝑀 × 𝑁 data
𝑙=𝐿1 𝑓 (𝑛, 𝑙)𝑆(𝑛, 𝑙) matrix of the form:
𝐹 (𝑛) = ∑𝐿2 (3)
[ ]𝑇
𝑙=𝐿1 𝑆(𝑛, 𝑙) X = x(1) x(2) ... x(𝑁 − 2) x(𝑁 − 1) (4)
where 𝐿1 and 𝐿2 are the FFT indices of the boundary of the [ ]𝑇
averaging region and 𝑓 (𝑛, 𝑙) and 𝑆(𝑛, 𝑙) are the frequency where x(𝑛) = 𝑥(𝑛) 𝑥(𝑛 + 1) ... 𝑥(𝑛 + 𝑀 − 1) .
and energy in the 𝑙𝑡ℎ frequency bin of the 𝑛𝑡ℎ time frame. An estimate of the correlation matrix R̂𝑥 can be com-
puted by R̂𝑥 = 𝑁1 X𝐻 X, where the superscript 𝐻 denotes
TR-IAA: A class of non-parametric frequency estimation the Hermitian of a matrix. Eigenanalysis is carried on R̂𝑥
methods most recently developed is based on the time re- to find the vectors spanning the signal and noise subspaces.
cursive iterative adaptive approach (TR-IAA) [7]. This algo- These subspaces are orthogonal to each other. The eigenvec-
rithm reaches the spectral estimates of a given time frame by tors q1 , q2 , ..q𝑃 corresponding to the largest 𝑃 eigenvalues
minimizing a quadratic cost function using a weighted least that span the signal subspace and the remaining eigenvectors
squares formulation. TR-IAA is an iterative technique, which (q𝑃 +1 , q𝑃 +2 , ..q𝑀 ) span the noise subspace.
takes from 10 to 15 iterations to converge, where the spectral MUSIC makes use of the orthogonality property of noise
estimate is initialized to be either that of the spectrogram or eigenvectors and steering vectors v(𝑓𝑘 ), 1 ≤ 𝑘 ≤ 𝑃 (corre-
the final spectral estimate of the preceding time frame [7] sponding to the actual frequency components of the signal).
[9]. This method is more computationally extensive than Here v(𝑓 ) is defined as:
spectrogram-based techniques. After the convergence of the
spectral estimate, a quadratic interpolation scheme similar [ ]𝑇
𝑣(𝑓 ) = 1 𝑒𝑗2𝜋𝑓 𝑒𝑗4𝜋𝑓 ... 𝑒𝑗2(𝑀 −1)𝜋𝑓 (5)
to the first spectrogram-based approach discussed is used to
estimate the frequency. ¯ 𝑚𝑢𝑠𝑖𝑐 , can then be com-
The pseudo-spectrum of MUSIC, 𝑅
It has been shown that the TR-IAA based approach pro- puted as:
vides slightly better frequency estimates than the spectrogram
based approach, when the frame size is 20-30 seconds [7]. ¯ 𝑚𝑢𝑠𝑖𝑐 (𝑒𝑗2𝜋𝑓 ) = ∑ 1 1
𝑅 𝑀
= 𝑗2𝜋𝑓 )∣2
Since we are considering short frame sizes and due to the 𝑚=𝑝+1 ∣v𝐻 (𝑓 )q𝑚 ∣2 ∣𝑄 𝑚 (𝑒
high computational costs of the TR-IAA based algorithm ob- (6)
served, we focus on spectrogram-based approaches for non- Due to the orthogonality property, the denominator should be
parametric methods in this paper. zero at the frequencies of the signal.
Among various techniques studied in the statistical signal
2.2. Parametric Approaches processing literature, RootMUSIC provides high precision at
a moderate computational cost. It solves for the roots of the
Parametric methods assume an explicit model for the signal denominator directly [10]. The frequency estimates using this
and underlying noise. Due to such an explicit assumption method are the arguments of the 𝑃 roots closest to the unit
about the model, the estimates obtained using parametric ap- circle. We opted to use RootMUSIC in our comparisons.
proaches are more accurate than non-parametric approaches
[10]. In this paper, we consider two of the most widely used ESPRIT: ESPRIT makes use of the rotational property be-
parametric frequency estimation methods based on the sub- tween staggered subspaces that is invoked to produce the fre-
space analysis of a signal and noise model. In general, these quency estimates. In our case, this property relies on observa-
methods can be used to estimate the frequency of a signal tions of the signal over two intervals of the same length stag-
composed of 𝑃 complex frequency sinusoids embedded in gered in time. ESPRIT is similar to MUSIC in the sense that
they are both subspace-based approaches, but it is different 3. COMPARISON OF FREQUENCY ESTIMATION
in that it works with the signal subspace rather than the noise METHODS
subspace. The implementation that we employ estimates the
signal subspace from the data matrix X of Eq. 4. In this section, we compare the performance of different
A Singular Value Decomposition (SVD) is applied to X, frequency estimation methods discussed in Section 2. We
giving: present the experiments that we carry out and the criteria by
which we compare the methods.
X = LSU𝐻 (7)

where L is an 𝑁 × 𝑁 matrix of the left singular vectors, S is 3.1. Experiments on Synthetic Signals
an 𝑁 × 𝑀 matrix with its main diagonal entries containing
the singular values, and U is an 𝑀 × 𝑀 matrix of the right To facilitate a comparative study, we first generate synthetic
singular vectors. The singular values correspond to the square signals with ground truth frequencies. In our model of the
roots of the eigenvalues of the sample correlation matrix R̂𝑥 synthetic signal, the frequency of the signal changes on a
scaled by N, and the columns of U are the eigenvectors of frame basis and the frequency estimation algorithms are ap-
R̂𝑥 . These vectors form an orthonormal basis for the under- plied to each time-frame. Since the ENF signal has a slowly-
lying 𝑀 -dimensional vector space. More specifically, U can varying frequency, we attempt to capture the correlation of the
be written as U = [U𝑠 ∣U𝑛 ], where U𝑠 is the 𝑀 × 𝑃 matrix frequencies in consecutive frames by first generating a ran-
of right singular vectors corresponding to the singular values dom sequence of frequencies having a normal distribution of
with the 𝑃 largest magnitudes and U𝑛 is the 𝑀 × (𝑀 − 𝑃 ) mean 𝜇 = 60 and standard deviation 𝜎 = 0.0133. Then, to
matrix containing the remaining right singular vector. The mimic the pseudo-periodic behavior of the ENF fluctuations,
signal subspace can be partitioned into two smaller (𝑀 − 1)- we pass this sequence through a filter of the form [13]:
dimensional subspaces as: 1
𝐻(𝑧) = . (11)
[ ] [ ] 1 − 0.97𝑧 −1
U1 ∗
U𝑠 = = (8) For one instance of this simulation, the synthetic signal is gen-
∗ U2 ,
erated as a series of consecutive sinusoidal signals of frequen-
cies corresponding to the generated sequence, each with a
where U1 and U2 correspond to the unstaggered and stag-
phase following a uniform distribution 𝑈 (0, 2𝜋). We consider
gered subspaces, respectively. The relation between U1 and
the sampling frequency of 441Hz, and examine several frame
U2 can written as:
sizes, ranging from 0.1 seconds (44 samples) to 1 second (441
U2 = U1 Q, (9) samples). We add additive white gaussian noise (AWGN) to
where Q is a 𝑃 × 𝑃 matrix. Q can be computed using a achieve different levels of signal-to-noise ratio (SNR).
least squares method. Eigenanalysis can then be carried out We carried out 100 simulation runs where each run has
on Q. The frequency estimates can be extracted from the ar- a different set of frequencies, phase angles, and additive
guments of the eigenvalues of Q. Denoting these eigenvalues noise. In each simulation run, the frequencies are estimated
by 𝜙𝑘 , 1 ≤ 𝑘 ≤ 𝑃 , we can find the frequency estimates, 𝑓ˆ𝑘 , on a frame by frame basis using the frequency extraction
by: methods discussed in Section 2. For the spectrogram-based
∠𝜙𝑘 approaches, the number of FFT points is chosen such that
𝑓ˆ𝑘 = with 1 ≤ 𝑘 ≤ 𝑃. (10) the frequency resolution is approximately 0.03 Hz. For the
2𝜋
weighted energy approach, the range considered for weight-
It is worth noting that the accuracy of the estimates using ing is [59.8, 60.2] Hz. For the subspace methods, we use the
subspace methods can differ significantly depending on the experimentally optimized 𝑀 value for the dimension of the
parameter 𝑀 chosen for the data matrix in Equation 4. It was data matrix based on the discussions in Section 2.2.
shown in [11] that for estimating the frequency in a single A direct way of comparing the frequency estimates is to
sinusoid in white noise, the error variance is minimal when 𝑀 subtract each sequence of estimates from the true sequence
is either 𝑁3 or 2𝑁3 . For a general multiple signal case, a rule of frequencies and estimate the mean difference. This crite-
of thumb suggested in [12] is that 𝑀 should be in the range rion does not fit our application because it may penalize some
[ 2𝑁 3𝑁
5 , 5 ]. So, before we carry out the experiments detailed in methods that tend to have an inherent bias in estimating the
the next section, we test the accuracy of MUSIC and ESPRIT frequency, such as the weighted energy approach. Since we
on the range 𝑀 ∈ [ 𝑁3 , 2𝑁 3 ] for the 𝑁 that we intend to use. are interested in measuring the similarity in the trends of the
We find the value of 𝑀 that gives the most accurate result signals for most applications based on ENF signal analysis,
for each case of 𝑁 , and subsequently use it in the following these trends are better revealed using a correlation based met-
experiments. The metric used for determining accuracy is the ric.
correlation coefficient, which will be discussed in the next More specifically, to measure the performance of different
section. frequency estimation methods, we use the cross-correlation
 0.4 1

ESPRIT
0.35 0.9
MUSIC
Spectrogram − Quadratic Fitting
0.3 Spectrogram − Weighted Energy
0.8
Correlation Coefficient

Correlation Coefficient
0.25
0.7
0.2
0.6 ESPRIT
0.15 MUSIC
Spectrogram − Quadratic Fitting
0.5
0.1 Spectrogram − Weighted Energy

0.05 0.4

0
0 5 10 15 20 0 5 10 15 20
SNR (dB) SNR (dB)

(a) 0.1-second frame (b) 0.5-second frame

0.95

0.9
Correlation Coefficient

ESPRIT
0.85 MUSIC
Spectrogram − Quadratic Fitting
Spectrogram − Weighted Energy
0.8

0.75

0.7

0.65
0 5 10 15 20
SNR (dB)

(c) 1-second frame

Fig. 1. Synthetic Signals: Correlation Coefficient vs. SNR for very short frame sizes

between the frequency estimates and the true frequencies and and this signal model matches the synthetic model well. As
measure the correlation for different frame sizes and different with ESPRIT outperforming MUSIC, our observation in the
SNR values. The results are shown in Fig. 1. The correla- context of the ENF problem is consistent with the general re-
tion coefficient of two sequences that do not match is close to sult in the literature that ESPRIT gives slightly more accurate
zero, and the correlation coefficient for matching sequences estimates than MUSIC for a similar computational cost [14].
is expected to be higher. From Fig. 1, we can see that the fre-
quency estimates worsen when more noise is added and the
3.2. Experiments on Power Grid Data Set
SNR decreases, which is expected. Also, when the frame size
becomes small, as with Fig. 1(a), all estimation techniques To validate the results that we have obtained using synthetic
behave rather poorly, as the correlation coefficient achieved data, we examine the performance on a power signal mea-
was around 0.35 in high SNR cases and only 0.05 in low SNR surement available at [15]. This data set also provides a se-
cases; the latter is barely differentiable from matching two un- quence of reference frequencies computed using a Frequency
related ENF signals. For this reason, it is advisable not to use Disturbance Recorder (FDR). FDR estimates the frequencies
such small frame sizes whenever possible. using phasor analysis and signal resampling techniques, and
We also note that both spectrogram-based approaches the FDR measurement is known to provide a resolution of
give similar performances; the subspace-based approaches around ±0.0005 Hz [16].
give better performances than the spectrogram-based ap- We assume the power measurement to be noise-less for
proaches, with ESPRIT consistently outperforming MUSIC simplicity, and we add various levels of AWGN to it. We es-
by a moderate margin. This is due to an explicit assumption timate the frequencies present for different frame sizes using
on the sinusoidal signal model in the parametric approaches the estimation techniques discussed in Section 2. We use the
 0.9
Table 1. Correlation coefficient values between the ENF sig-
0.8 ESPRIT nals extracted from the power and audio recordings.
MUSIC
Spectrogram − Quadratic Fitting
0.7
Spectrogram − Weighted Energy Method Correlation Coefficient
Correlation Coefficient

0.6
ESPRIT 0.9778
0.5 MUSIC 0.9767
Spectro. – Quad. Interp. 0.9704
0.4
Spectro. – Weighted Energy 0.9543
0.3

0.2

0.1
0 5 10 15 20 quency, to remove as much noise as possible without affecting
SNR (dB)
the frequency band of interest. We partition both the power
signal and the filtered audio signals into overlapping frames,
Fig. 2. Power Grid Data Set: Correlation Coefficient vs. SNR and compute the frequency estimates for each frame using
for frame size of 1 second the four ENF extraction techniques under study. In order to
differentiate the various estimation methods, we compute the
correlation coefficient between the audio and power estimates
frequencies computed by the FDR as a reference, and com-
of each method.
pute the correlation coefficient between this reference and the
estimated sequence of frequencies. This correlation was com- The results are shown in Table 1, where a higher value of
puted after temporally aligning the reference frequency se- the correlation coefficient suggests better matching between
quence with the estimated ones. the ENFs extracted from the audio and power signals. Al-
though the values in Table 1 are close, they support the find-
The results for 1-second frame size can be seen in Fig. 2.
ings reached earlier that ESPRIT gives the best results fol-
The estimates that we obtained are worse than the estimates
lowed by MUSIC. This table also demonstrates to us that the
for synthetic data, which is understandable as the model we
spectrogram-based quadratic interpolation approach performs
used was an idealization of the reality and the power measure-
better than the weighted energy approach. This can be due
ments is likely not noise free or a perfect sinusoid. However,
to our restricting the estimates of the weighted energy to a
the order of accuracy remains the same as with the synthetic
chosen range which may not be true in all signal cases. The
data comparison, with ESPRIT outperforming the other three
quadratic interpolation approach has no such restraints.
methods, followed by MUSIC. Fig. 2 shows similar trend and
relative positions to those shown in Fig. 1(c) except that the
correlation coefficient values are lower. 4. ENF SIGNAL AS A LOCATION STAMP

3.3. Matching ENF Signals from Audio and Power We have demonstrated in Section 3 the performance of dif-
ferent frequency estimation algorithms for ENF analysis us-
A main forensic application involving ENF analysis requires ing a correlation based metric. In this section, we explore
matching an ENF signal extracted from an audio signal to a novel application of ENF signal analysis for spatial local-
that extracted from a power signal in order to determine the ization of sensing signals. The ENF signal has been used to
time of recording. Here, we carry out experiments to compare determine the time of recording of a given multimedia sig-
which of the four estimation methods gives the best matching nal that contains traces of ENF variations during the sensing
between ENF signals that are known to be recorded at the process [1] [2]. Coarse location information in terms of the
same time. grid in which the recording was made may also be determined
More specifically, we make two simultaneous recordings using ENF analysis. The specific ENF variations are not the
of audio and power signals. The power signal is obtained same for recordings made over different power grids that are
from an electric outlet using a step-down transformer. The not frequency synchronized, for example, in the east intercon-
audio signal is obtained by recording the background noises nection grid versus the west interconnection grid. Different
in a room. As mentioned in the introduction section, the audio degrees of controls in regulating the power grid also play a
recording may pick up the ENF signal from electromagnetic key role [4]. The range of fluctuations in the ENF signal also
interferences in the room. We expect that if there is ENF generally depends on the size of the grid, as a smaller capac-
present in the audio, it should match with that in the power ity grid often exhibits a higher range of fluctuations, while a
recording captured concurrently. large capacity grid can be more tightly controlled and exhibits
Prior to ENF extraction, we pass the audio signal through a smaller range of fluctuations. For example, ENF variations
a bandpass filter centered around 60Hz, the nominal fre- in the UK power grid are over a higher range than the conti-
nental European Interconnection grid [4]. 60.06

60.05

60.04

Frequency (in Hz)

60.03

60.02
MD1
60.01 MD2
NJ1
60
NJ2
59.99 MA1
MA2
59.98
0 200 400 600 800 1000
Time (in seconds)

Fig. 4. Aligned ENF signals recorded in MD, NJ and MA

(best viewed in color)

of the ENF signal propagation across the grid can be poten-

tially used for localization at a finer resolution within a grid by
Fig. 3. Map of the locations where ENF recordings were con- comparing the similarity of the ENF signal in question with
ducted. [Source of base map: www.d-maps.com] the ENF databases that may be available for a set of locations.
Devising a methodology to pinpoint the place of record-
In addition to inter-grid variations, ENF signals may also ing is useful for forensic analysis on the media and sensor
exhibit intra-grid variations. These variations among ENF recordings although locating within a grid using ENF traces
signals within the same grid are due to the local load charac- has not been demonstrated in the literature to the best of our
teristics of a given city and the time taken to propagate the re- knowledge. In this section, we explore the presence of lo-
sponse to demand and supply in load to other parts of the grid cation specific signatures based on the experiments on ENF
[4]. In general, power demand and supply in a given area fol- signals obtained from three locations in the Eastern intercon-
lows a cyclic pattern; for example, demand increases during nection grid of the United States.
evening hours in a residential neighborhood. Any change in
the load in a grid is regulated by a control mechanism [4]. An 4.1. Experiments Conducted and Observations
increase in the load causes the supply frequency to drop tem-
porarily; the control mechanism senses the frequency drop We record power signals in Maryland (MD), New Jersey (NJ),
and starts drawing power from adjoining areas to compensate and Massachusetts (MA) in the United States. These states
for the increased demand. As a result, the load in adjoining are part of the Eastern interconnection grid. As shown in
areas also increases, which leads to a drop in the instanta- the map of Fig. 3, the recordings are conducted in College
neous supply frequency, and the overall power supply will Park in MD, Princeton in NJ, and Cambridge in MA. The
be driven up to compensate the rising load which leads to a approximate distance between College Park and Princeton is
drop in the instantaneous supply frequency in those regions. 153 miles, while the distance between Cambridge and Prince-
A similar mechanism is used to compensate for an excess sup- ton is 232 miles. Power signals are recorded by providing a
ply of power flow that leads to a surge in supply frequency. step-down voltage from power-mains supply as an input to the
A small change in the load in a given area may have a sound card of Olympus WS700 series audio recorders. The
localized effect on ENF in the given area. However, a large sampling frequency of the Olympus recorders is 44.1 kHz,
change, such as one caused by a generator failure may have an and it does not provide alternative sampling rates. At each lo-
effect on the whole grid. These changes are known to prop- cation, we conduct two separate recordings at the same time
agate along the grid at a typical speed of approximately 500 at close vicinity to each other. We record power signals from
miles per seconds [5]. We conjecture that small and large these locations for multiple days. All recordings are started
changes in the load may cause location specific signatures in simultaneously by synchronizing with a webclock [17].
the ENF patterns, and if so, such differences may be exploited Each recorded power signal is first lowpass filtered and
to narrow down the place of a given recording within a power downsampled to 250 Hz. A bandpass filter with passband in
grid. Due to a finite speed of propagation of frequency dis- 59-61 Hz is then used to extract the ENF signal at 60 Hz. In-
turbances across the grid, we anticipate that the ENF signal stantaneous frequency is estimated using the ESPRIT method
would be more similar for locations close to each other as by dividing the signal into 2-seconds long frames, and each
compared with the locations farther apart. Such a property frame has a 1-second overlap with the previous frame. Al-
 1 60.05
NJ1−NJ2 MD
0.9998 MA1−MA2 NJ
MD1−MD2 60.045 MA
Correlation coefficient

0.9996 MD2−NJ2

Frequency (in Hz)

MD2−NJ1
0.9994 60.04
NJ2−MA2
NJ1−MA2
0.9992
MD2−MA2 60.035
0.999 MD1−MA2

60.03
0.9988

0.9986 60.025
0 10 20 30 40 180 200 220 240
segment index Time (in seconds)

(a) Correlation coefficient

Fig. 6. Zoomed view of the ENF signals in different states
0.8 (best viewed in color)
MD1−MD2
0.7 NJ1−NJ2
MA1−MA2
0.6
MD2−NJ1 ever, a careful “microscopic” examination reveals some dif-
MD2−NJ2
0.5 ferences in the ENF sequences. A zoomed plot of an ENF
Histogram

NJ1−MA2
0.4 NJ2−MA2 sequence from a recording from each of the three locations is
0.3
MD1−MA2 shown in Fig. 6. From this figure, we observe that the ENF se-
MD2−MA2
quences at each location have different location specific traits,
0.2
which are not present at other locations. Location-specific
0.1 traits are present in the recordings conducted across all three
0 cities. Next, we explore such location-specific signatures.
0.992 0.994 0.996 0.998 1
Correlation coefficient

(b) Histogram of Correlation coefficient

4.2. Exploring Location Signatures
Fig. 5. Correlation coefficient between ENF signals ob- In this subsection, we describe a method used to extract the
tained across different locations for 500-second query seg- location-specific characteristics from each recording. As
ment length. shown in Fig. 6 earlier, the microscopic trends in the ENF
signal recorded at distinct locations are different. These de-
tailed trends vary fast with time, and are present at higher
though the recordings were manually synchronized using a frequency details in the ENF signal. This motivates us to
web clock at [17], misalignment of the ENF sequences may apply a high-pass filter to extract such location specific sig-
still be present. As the main trends of ENF signals are known natures. To realize such a high-pass filter, we pass each ENF
to be the same at locations within a grid, we temporarily align signal through a smoothening filter, and subtract the resulting
all ENF sequences by finding the time corresponding to the signal from the original ENF sequence. The highpass filtered
maximum value of their pairwise correlation coefficient. Fig. ENF signal, 𝑓ℎ𝑝 (𝑛), can be written as:
4 shows a plot of a 1000 samples of aligned ENF sequences
𝑁 −1
recorded on 20-21 July 2012 across the three locations de- ∑ 2

scribed above. From this figure, we observe that as expected, 𝑓ℎ𝑝 (𝑛) = 𝑓 (𝑛) − 𝑤(𝑘)𝑓 (𝑛 − 𝑘) (12)
the ENF signals at a given time are similar at different loca- 𝑘=− 𝑁 2−1
tions across the same grid. We compute the pairwise corre-
lation coefficient between the recorded ENF signals for seg- where 𝑓 (𝑛) is the ENF value at time 𝑛, 𝑤(⋅) is the coefficient
ments of 500-samples long. Each sample is the frequency of the smoothening filter, and 𝑁 is the order of the filter. We
estimate of a 2-second frame, with 1-second overlap with the choose an odd order for the filter.
previous frame, as explained earlier. We plot these correla- After obtaining the high-pass filtered ENF sequence for
tion coefficients in Fig. 5(a) and its histogram plot is shown each recording, we use the correlation coefficient of the high-
in Fig. 5(b). From these figures, we observe that the ENF se- pass filtered ENF sequences as a metric to examine the loca-
quences across different cities in the same grid recorded at the tion similarity between two recordings. If the proposed filter
same time are highly correlated with each other. is able to capture the location specific signatures, we expect
It is difficult to distinguish the signals recorded at different to obtain a high value of correlation coefficient for intra-city
locations using such a macroscopic correlation plot. How- recordings, and a low-value for inter-city recordings.
 the pairwise correlation coefficient between intra-city record-
1 MA1−MA2 ings is close to 1, and for inter-city recordings, it is much less
NJ1−NJ2
0.8 MD1−MD2 than 1. This observation suggests that the highpass filter can
Corrleation coefficient

MD2−NJ2 potentially capture the city specific signatures from the ENF
MD2−NJ1
0.6 signal.
NJ1−MA2
NJ2−MA2 Among the inter-city data, we observe that the value of
0.4
MD1−MA2 correlation coefficient for MD-NJ and NJ-MA recordings are
MD2−MA2
0.2 higher than those for MD-MA recordings. As can be seen
earlier in Fig. 3, Princeton, NJ, is located approximately mid-
0
way between College Park, MD, and Cambridge, MA, and so
its distance to College Park and Cambridge is smaller than the
5 10 15 20 25 30 35 40
segment index distance between College Park and Cambridge. A high value
of correlation coefficient for MD-NJ and NJ-MA recordings
(a) Correlation coefficient
as compared to the correlation coefficient between MD-MA
0.7 recordings reflects the time delay from the propagation effect
MD1−MD2 of the load balancing, as was explained in Section 4. We
0.6 NJ1−NJ2
MA1−MA2 observe from Fig. 7(a) that the values of correlation coeffi-
0.5 MD2−NJ1 cient for MD-NJ recordings are moderately higher than NJ-
MD2−NJ2 MA recordings. This reflects the smaller distance between
Histogram

0.4 NJ1−MA2
NJ2−MA2 College Park and Princeton as compared to the distance be-
0.3 MD1−MA2 tween Princeton and Cambridge as shown in Fig. 3. These
MD2−MA2 plots demonstrate that the ENF signal has a strong potential
0.2
to be used as a location stamp by comparing the correlation
0.1 coefficient of the highpass filtered query ENF signal with the
0 highpass filtered ENF signal database of each city from the
−0.4 −0.2 0 0.2 0.4 0.6 0.8 1
Correlation coefficient corresponding time.
(b) Histogram of correlation coefficient
From the three-city data for the U.S. east coast, we ob-
serve that the correlation coefficient between the highpass fil-
tered query ENF signals captured across different cities in a
Fig. 7. Correlation coefficient between ENF signals across
geographical area covered by a grid is approximately pro-
different locations for 500-second query segment length when
portional to the distance between the locations. As a result,
high-pass filtered ENF signal is used for location matching.
the correlation coefficient can be used as an indicator of the
distance between two cities. Given the distance estimated
4.3. Results and Discussions between cities and the known location-of-recording of some
ENF signals behaving as anchor signals, trilateration tech-
We now discuss the results obtained from a 14-hour long niques may be used to estimate the recording location of a
recording of the power ENF signal conducted across the query. More generally, however, it may become difficult to
three different locations described above, namely, College characterize a simple relationship between correlation coef-
Park, Princeton, and Cambridge/Boston. After aligning the ficients and distances, as the correlation coefficient may de-
estimated ENF sequences temporally, we pass each signal pend on such factors as power line density in a grid, and wired
through a smoothening filter of order 𝑁 = 3. We use a uni- distances may vary from the flight or road distances between
form weight vector 𝑤(𝑘) = 𝑁1 for all 𝑘. This is followed the cities. To mitigate this issue, we are collecting data from
by subtracting the smoothed version from the original ENF multiple locations in the eastern and the western grid of the
signal to obtain a highpass filtered ENF signal. We divide United States to empirically characterize correlation coeffi-
the high-pass filtered signals into non-overlapping segments cient profiles of the grid with geographical distance. Such
of 500 samples each. The correlation coefficient of each a study will help in a better understanding of the resolution
segment for the recordings across different cities is shown limits of the location stamping properties of the ENF signals.
in Fig. 7(a). From this figure, we observe relatively higher
values of correlation coefficient for all intra-city recordings 5. CONCLUSION AND FUTURE WORK
(MD1-MD2, NJ1-NJ2, MA1-MA2) for the top three curves
as compared with inter-city recordings. In this paper, we have studied different parametric and non-
We further plot the histogram of correlation coefficients parametric frequency estimation methods with application to
between different highpass filtered ENF signals across differ- Electric Network Frequency (ENF) signal estimation in high
ent locations in Fig. 7(b). From this figure, we observe that temporal and frequency resolution. We have conducted ex-
periments on synthetic data and experimental data under dif- [8] J. O. Smith and X. Serra, “PARSHL: An analysis/synthesis
ferent noise conditions to evaluate the performance of stud- program for non-harmonic sounds based on a sinusoidal repre-
ied methods using a correlation coefficient based metric. Our sentation,” International Computer Music Conference, 2004.
results have demonstrated that the ESPRIT based paramet- [9] G. O. Glentis and A. Jokobssen, “Time-recursive IAA spectral
ric frequency estimation method provides the best results for estimation,” IEEE Signal Processing Letters, vol. 18, no. 2,
ENF matching using correlation coefficient. We have also Feb. 2011.
studied a novel application of ENF signal analysis in pin- [10] D.G. Manolakis, V. K. Ingle, and S. M. Kogon, Statistical and
pointing locations based on the characteristics of ENF signals Adaptive Signal Processing, MaGraw-Hill, Inc., 2000.
recorded in several cities of the United States Eastern inter- [11] A. C. Kot, S. Parthasarathy, D. W. Dufts, and Y. Ding, “Statis-
connection power grid. We have proposed a method based tical performance of single frequency estimation in white noise
on high-pass filtering to extract location traces in the ENF using state variable balancing and linear prediction,” IEEE
signal. Preliminary results of the proposed approach demon- Transactions on Acoustics, Speech, and Signal Processing, vol.
strate that the ENF signal carries location specific traces that 35, pp. 1639–1642, Nov. 1987.
have a strong potential towards estimating the location of me- [12] Y. Ding and R. J. Vaccar, “Determination of data matrix dimen-
dia and sensor recordings. sions for subspace-based parameter estimation algorithms,”
Our ongoing and future work includes conducting a more IEEE International Conference on Acoustics, Speech, and Sig-
comprehensive study of ENF signal based location-stamping nal Processing, vol. 5, pp. 2547 – 2550, May 1996.
by collecting more ENF data across different power-grids in [13] R. Garg, A. L. Varna, and M. Wu, “Modeling and analysis
different cities and different sensing modalities. of electric network frequency signal for timestamp verifica-
tion,” To appear in IEEE International Workshop on Infor-
mation Forensics and Security, Dec. 2012.
Acknowledgment [14] P. Stoica and R. Moses, Spectral Analysis of Signals, Prentice
Hall, Inc., 2005.
The authors would like to thank Dr. Shan He, Mr. Chau-Wai
Wong, and Mr. Michael Luo for their help in the power data [15] “http://www.sal.ufl.edu/newcomers/enf data.rar,” Accessed
collection for the research work done in this paper. July 2012.
[16] Y. Zhang, P. Markham, T Xia, L. Chen, Y. Ye, Z. Wu, Z. Yuan,
6. REFERENCES L. Wang, J. Bank, J. Burgett, R. W. Conners, and Y. Liu,
“Wide-area frequency monitoring network (FNET) architec-
[1] C. Grigoras, “Applications of ENF criterion in forensics: Au- ture and applications,” IEEE Transactions on Smart Grid, vol.
dio, video, computer and telecommunication analysis,” Foren- 1, no. 2, Sep. 2010.
sic Science International, vol. 167, no. 2-3, pp. 136 – 145, Apr. [17] Web-Based Clock, “http://www.timeanddate.com,” Accessed
2007. July 2012.
[2] R. Garg, A. L. Varna, and M. Wu, “‘Seeing’ ENF: natural time
stamp for digital video via optical sensing and signal process-
ing,” in Proeedings. of the 19th ACM International Conference
on Multimedia, 2011, pp. 23–32.
[3] R. Garg W. H. Chuang and M. Wu, “How secure are power net-
work signature based time stamps?,” ACM Conf. on Computer
and Communication Security, Oct. 2012.
[4] M. Bollen and I. Gu, Signal Processing of Power Quality Dis-
turbances, Wiley-IEEE Press, 2006.
[5] S. J. Tsai, L. Zhang, A.G. Phadke, Y. Liu, M.R. Ingram, S.C.
Bell, I.S. Grant, D.T. Bradshaw, D. Lubkeman, and L. Tang,
“Frequency sensitivity and electromechanical propagation sim-
ulation study in large power systems,” IEEE Transactions on
Circuits and Systems, vol. 54, no. 8, pp. 1819 –1828, Aug.
2007.
[6] S. Massoud Amin and B.F. Wollenberg, “Toward a smart grid:
power delivery for the 21st century,” IEEE Power and Energy
Magazine, vol. 3, no. 5, Sep. 2005.
[7] O. Ojowu, J. Karlsson, J. Li, and Y. Liu, “ENF extraction from
digital recordings using adaptive techniques and frequency
tracking,” IEEE Transactions on Information Forensics and
Security, vol. 7, no. 4, pp. 1330 – 1338, August 2012.