Detection of Forgery in Digital Video
Prof. Asok De
Dept. of Electronics and Communication Engineering
Delhi College of Engineering
University of Delhi

Himanshu Chadha
Dept. of Computer Engineering
Delhi College of Engineering
University of Delhi

Sparsh Gupta
Delhi College of Engineering
University of Delhi


ABSTRACT

In this paper we present a novel and independent way
to uncover forgery in Digital Videos using the Readout
Noise that is introduced into a frame upon the Readout
from the CCD of a digital camera. The use of Readout
Noise is essential to gain independence since the noise
introduced is characteristic of the camera. The Readout
Noise is extracted from the camera under consideration
by taking bias frames and doing the frame analysis. This
could also be done by averaging the noise obtained from
multiple frames. For each frame in a video we first
calculate the reference readout noise pattern and using
this, we try to find certain regions inside the frame
which show marked variation from the reference
calculated. Further, a comparison with the noise pattern
in successive frames confirms the integrity of the video.
The work is supported by forging videos using different
cameras and detecting the presence of forged areas.

Keywords: Readout Noise, Bias Frame. Frame Analysis.


INTRODUCTION

Digital counterparts are fast replacing the analog
cameras in our day to day life. The trend stands justified
keeping in mind the fact that digital files can be
conveniently shared over computer networks and may
be converted easily from one to the desired format.
Further, the ease of storage, and the convenience with
which these may be retained without the loss of quality
over time and usage, have made them being heavily
relied on.
But, with the ease of availability of powerful editing
programs today, it has become increasingly simple for
even amateurs to create real looking forgeries in images
and videos.
Hence, the confirming the integrity of digital image/
video is becoming a matter of concern with many
associated agencies.

In this paper we discuss a new approach to detect
forgery in digital videos by using the readout noise
which is, involuntarily, inserted into an image/frame
upon the readout from the CCD of a digital camera. The
readout noise pattern is observed over consecutive
frames to detect any marked variation. The variation is
calculated in reference to the noise pattern observed.
This reference noise pattern is calculated by denoising
the frames using a filter and then subtracting the
denoised signal from the original signal to obtain the
noise signal. The Readout noise for a given camera CCD
may also be calculated by doing the bias frame analysis.


READOUT NOISE
At its most basic level, just like a conventional camera, a
digital camera has a series of lenses that focus light to
create an image of a scene. But instead of focusing this
light onto a piece of film, it focuses it onto a
semiconductor device that records light electronically. A
computer then breaks this electronic information down
into digital data.
Instead of film, a digital camera has a sensor that
converts light into electrical charges. The image sensor

employed by most digital cameras is a charge coupled
device (CCD). Hence a CCD is the heart of a digital
camera.
Primary colors are identified by placing a filter called a
color filter array over each individual photosite. By
breaking up the sensor into a variety of red, blue and
green pixels, it is possible to get enough information in
the general vicinity of each sensor to make very
accurate guesses about the true color at that location.
This process of looking at the other pixels in the
neighborhood of a sensor and making an educated
guess is called interpolation.
Following is the order of operations the CCD performs
to turn the electrons knocked free by photons in every
pixel into a signal which can be read by a computer:
1. Electrons transferred to "amplifier"; really a
capacitor.
2. The voltage induced by this charge is measured.
3. An Analog-To-Digital (A/D) unit converts the
voltage into some other voltage, which may
have only one of several discrete levels.
4. The voltage is converted into a number which is
passed from the hardware to the computer
software as the pixel's value. Units are counts,
also called "Data Numbers" (DN) or "Analog-to-
Digital Units" (ADUs).
Readout noise occurs in step 2: the measurement of a
very small packet of charge by the readout amplifier.
Readout noise is the number of electrons introduced per
pixel into the final signal upon the readout of the device.
It consists in two components:
i. the A/D conversion is not perfectly repeatable
even for the hypothetical case of reading out
the same pixel twice and each time with the
same charge, the output value could be slightly
different;
ii. the electronics of the CCD introduce spurious
electrons into the process with unwanted
random noise in the output.
The presence of these two random effects, produce an
uncertainty in the final output value for each pixel. In
the output of the CCD image, readout noise is added
into every pixel every time the array is read out: this
means that a CCD with a readout noise of 20 electrons
will, on average, contain 20 extra electrons of charge in
each pixel upon readout.

FORGERY DETECTION ALGORITHM

Forgery in a video may be of two kinds. First, where an
entire frame is inserted into frames of a video and
second where a small area is forged and it appears in
one or many frames.

Forgery of the first kind may be detected by averaging
the noise over all the frames and then comparing the
mean noise with the noise in each and every frame to
detect forgery. The first step to detection is to calculate
the noise pattern for the frames of a video. For
calculating the noise in a frame, we apply a denoising
filter to the captured frame and subtract this frame from
the original frame to obtain the noise pattern.

Let the kth frame of a video be denoted by and let
be the denoising filter [2]. Then the noise pattern
for frame is given by:



The denoising filter [2] removes noise from the frame to
give a frame with only the actual capture information.
This frame is subtracted from the original frame to
obtain a frame with only pure noise. Hence a noise
pattern may be obtained for all the frames of a digital
video. This noise pattern obtained is averaged over all
the frames to obtain the mean noise.

However, mean noise may also be obtained for a given
camera by taking two dark frames [6] with a zero time
exposure. Let these frames be called Frame1 and
Frame2. We, now, add an offset value (o) to Frame2 so
that the subtraction of Frame1 from Frame2 does not
produce negative values. The output to this procedure is
a frame with pure noise averaged over the offset value
(o). The output is further divided by √2 since combining
images (like the subtraction above) always increases the
noise quadratically. [6]




Where, is the standard deviation of the
difference between two bias frames, RN represents

Readout Noise and G represents the Gain. While, B1, B2
and readout noise are expressed in e- the gain is
expressed in e-/ADU. The complete procedure,
however, has been explained clearly in [6].

The next step is to calculate the variance of the noise of
a particular frame with the mean noise. This may be
done by using the mathematical expression:



Where, N is the mean noise and represents the
variance of noise of kth frame:

This mean noise is used as a reference and variance is
found between the mean noise and the noise of a
particular frame. A frame with high variance from the
mean of the frame is a suspect frame that may not
belong to the video. However, a frame that does belong
to that specific video will show low variance from the
mean of the noise calculated.

Another kind of forgery generally found in digital videos
is when a part or a small area of a video frame is forged
using editing programs. This type of forgery may be
detected by dividing a video frame into small areas and
determining the correlation pattern between the noise
pattern in a specific area and the mean noise pattern
observed in the frame and the video as a whole. This
would help us to determine the areas inside a video
frame which show remarkable difference compared to
the results of the other areas in the same frame.

Let the noise pattern of an area be denoted by
where there are ‘m’ areas and 0<i<m+1. Let the frame
noise be denoted, as before, by . Let denote, as
before, the average camera noise. Then a correlation
between the noise in an area and noise in that frame
would help us detect the area missing the CCD readout
noise.




Where, is the noise in an area, is the average
noise of a fram e index ‘ ’ and denotes the average
noise of a camera i.e. the average of all frames or the
camera noise calculated by taking bias frames and doing
the frame analysis.
All areas in any frame that belong to a particular video
will show a general correlation pattern and the value of
the function, , will not vary beyond a
obtained threshold value. A forged area, however, will
show marked variation in this correlation pattern
leading to its identification. The forged area is easily
identified as this area is characterized by absence of the
unique readout noise associated with the CCD of a
camera.

The above said detection algorithm works well for the
case when a number of frames are forged, all at the
same area. We search for the area which does not
exhibit the same noise pattern as the other parts of the
frame.

Another kind of forgery, where an area in only one of
the frames is forged, is also possible. We were able to
develop a very efficient algorithm for this kind of
forgery.
If a video contains around 20 frames per second then
difference between noise patterns over two consecutive
frames would be small. We use this relation of the
consecutive frames to address the problem of forgery
detection.
Let a noise pattern of a frame be denoted by and
let denote the noise in an area. Let denote
the noise of ith area of kth frame. An average noise
pattern for an area then may be denoted by




Where, denotes the mean value of noise of ith area
in k
th
frame.

This evaluates the mean of an area noise over three
consecutive frames, one frame being before and one
being after a particular frame. Then, variance may be
calculated between the mean noise of an area to the
actual noise for all values of . The area showing high
variance value with respect to the mean noise
associated with it over consecutive frames is a
suspected forged area.



where symbols, used, hold their respective values. This
pattern would help us take advantage of the fps (frames
per second) attribute of a video, since a video with high

fps rate would have better correlation factor amongst
consecutive frames and detection of a forged area using
the above algorithm becomes efficient.

The complete detection algorithm developed by us is
listed below explicitly:

1. Obtain the video in frames. A video with higher
‘Fram e per Second’ rate m akes detection
easier.
2. Obtain the denoised frame for each
frame to be tested.
3. Subtract the denoised frame from the original
frame to obtain noise frame.
4. Obtain the mean noise for the noise frames
under consideration. Using this mean noise,
we, calculate the variance of each frame noise.

5.

6. Any noise frame with a variance which does not
match with the variance pattern observed over
many noise frames may be an inserted frame
which does not belong to the video.
7. Further, we divide a noise frame into small
areas and try to find the correlation between
the noise in an area and noise in that frame.

8.

9. Again a forged area will show marked variation
in the correlation pattern compared to the
integrated/original areas in a frame.
10. If a video contains more than 10 frames per
second then difference between noise patterns
over two consecutive frames would be small.
11. We,hence, calculate the noise in particular area
averaged over consecutive frames.

12.

13. Again an area showing high variance value with
respect to the mean noise associated with it
over consecutive frames is a suspected forged
area.

14.

15. Since, the type of forgery involved is not known
beforehand, hence, all the procedures
described in steps 4, 7 and 10 are repeated
over parts of video to observe forgery.



RESULTS AND CONCLUSION

In this paper, we have used a new approach to detect
the presence of forgery in digital video. This detection
may be carried out by using the readout noise pattern
for a specific camera. Two kinds of forgery have been
dealt with. The first where a video contains spurious
frames, in which case, the detection is carried out by
comparing the frame noise to average noise of the
video. Another kind of forgery where an area is forged in
many frames is detected by comparison of the noise
present in a specific area to the average noise of the
frame. A special case, where, an area is forged in only
one of the frames amongst consecutive frames is also
detected using a very efficient algorithm. The frames are
divided into areas and the noise in each area of a frame
is compared to the mean noise of that area, where, this
mean is calculated by averaging the noise over
consecutive frames.
Since, the kind of forgery carried out in a video is not
known beforehand hence we recommend the detection
to be implemented using multiple paths.

Since, the work is completely manual, our future plan is
to introduce automation such that the area division is
done efficiently, automatically and the detection may
not remain manual.

Forgery detection algorithm explained above was
implemented on forged video and difference in noise
pattern by about 3% was obtained for a case where a
video frame was forged using another video frame
captured by the sam e cam era ‘Creative PC- Cam 930
slim ’. This 3% difference was obtained between a forged
frame and an integrated/original frame of the video. The
results for the case where a different camera is used for
forgery, which is generally the situation, could not be
obtained due to the inavailability of another camera.
The difference in the noise pattern is expected to be
much larger for the case where video from a different
camera is used to forge a video from another source.

REFERENCES

1. Jan Luká sˇ, Jessica Fridrich, and M iroslav Goljan:
“Digital Camera Identification From Sensor Pattern
Noise” ,2005.
2. Mihcak M.K., Kozintsev, I., and Ramchandran, K.:
“Spatially adaptive statistical m odeling of w avelet
im age coefficients and its application to denoising,”
Proc. IEEE Int. Conf. Acoustics, Speech, and Signal
Processing, Phoenix, Arizona, vol. 6, pp. 3253–3256,
March 1999.
3. Popescu A.C. and Farid H.: “Exposing Digital
Forgeries in Color Filter Array Interpolated Im ages”,
IEEE Transactions on Signal Processing, vol. 53(10),
pp. 3948–3959, 2005.
4. Popescu A.C. and Farid, H.: “Exposing Digital
Forgeries by Detecting Duplicated Im age Regions”,
Technical Report, TR2004-515, Dartmouth College,
Computer Science 2004.
5. Cox, I., Miller, M.L., and Bloom, J.A.: Digital
Watermarking, Morgan Kaufmann, San Francisco,
2001.
6. “Readout Noise”.
http://users.libero.it/mnico/glossary/readnoise.htm
7. “Calculation of Readout Noise”.
http://users.libero.it/mnico/lab/tut_en7.htm
8. “Dark fram e subtraction,” Q im age Help [O nline].
http://www.ddisoftware.com/qimage/qimagehlp/dark.htm.
9. “Readout Noise, and Total Noise”.
http://spiff.rit.edu/classes/phys445/lectures/readout/readout.ht
ml.