Detecting Video Forgeries Based on Noise Characteristics.
ABSTRACT The recent development of video editing techniques enables us to create realistic synthesized videos. Therefore using video data as evidence in places such as a court of law requires a method to detect forged videos. In this paper we propose an approach to detect suspicious regions in video recorded from a static scene by using noise character- istics. The image signal contains irradiance-dependent noise where the relation between irradiance and noise depends on some parameters; they include inherent parameters of a camera such as quantum efficiency and a response function, and recording parameters such as exposure and elec- tric gain. Forged regions from another video camera taken under different conditions can be differentiated when the noise characteristics of the re- gions are inconsistent with the rest of the video.
- SourceAvailable from: Yoichi Sato[Show abstract] [Hide abstract]
ABSTRACT: Recently developed video editing techniques have enabled us to create realistic synthesized videos. Therefore, using video data as evidence in places such as courts of law requires a method to detect forged videos. In this study, we developed an approach to detect suspicious regions in a video of a static scene on the basis of the noise characteristics. The image signal contains irradiance-dependent noise the variance of which is described by a noise level function (NLF) as a function of irradiance. We introduce a probabilistic model providing the inference of an NLF that controls the characteristics of the noise at each pixel. Forged pixels in the regions clipped from another video camera can be differentiated by using maximum a posteriori estimation for the noise model when the NLFs of the regions are inconsistent with the rest of the video. We demonstrate the effectiveness of our proposed method by adapting it to videos recorded indoors and outdoors. The proposed method enables us to highly accurately evaluate the per-pixel authenticity of the given video, which achieves denser estimation than prior work based on block-level validation. In addition, the proposed method can be applied to various kinds of videos such as those contaminated by large noise and recorded with any scan formats, which limits the applicability of the existing methods.IEEE Transactions on Information Forensics and Security 01/2011; · 1.90 Impact Factor
- [Show abstract] [Hide abstract]
ABSTRACT: Digital forensics is one of the cornerstones to investigate criminal activities such as fraud, computer security breaches or the distribution of illegal content. The importance and relevance of this research fields attracted various research institutes leading to substantial progress in the area of digital investigations. One essential piece of evidence is multimedia data. For this reason this paper provides an overview of the state-of-the-art in the forensic investigation of multimedia data, the relationship between the various research fields and further potential research activities.IT Security Incident Management and IT Forensics, International Conference on. 01/2011;
- [Show abstract] [Hide abstract]
ABSTRACT: In this paper, a novel video inter-frame forgery detection scheme based on optical flow consistency is proposed. It is based on the finding that inter-frame forgery will disturb the optical flow consistency. This paper noticed the subtle difference between frame insertion and deletion, and proposed different detection schemes for them. A window based rough detection method and binary searching scheme are proposed to detect frame insertion forgery. Frame-to-frame optical flows and double adaptive thresholds are applied to detect frame deletion forgery. This paper not only detects video forgery, but also identifies the forgery model. Experiments show that our scheme achieves a good performance in identifying frame insertion and deletion model.Proceedings of the 11th international conference on Digital Forensics and Watermaking; 10/2012
Detecting Video Forgeries Based on Noise
Michihiro Kobayashi, Takahiro Okabe, and Yoichi Sato
Institute of Industrial Science, The University of Tokyo
Abstract. The recent development of video editing techniques enables
us to create realistic synthesized videos. Therefore using video data as
evidence in places such as a court of law requires a method to detect
forged videos. In this paper we propose an approach to detect suspicious
regions in video recorded from a static scene by using noise character-
istics. The image signal contains irradiance-dependent noise where the
relation between irradiance and noise depends on some parameters; they
include inherent parameters of a camera such as quantum efficiency and
a response function, and recording parameters such as exposure and elec-
tric gain. Forged regions from another video camera taken under different
conditions can be differentiated when the noise characteristics of the re-
gions are inconsistent with the rest of the video.
In the last decade digital cameras have become so popular that enormous num-
bers of photographs and videos are taken by amateur photographers. On the
other hand, the recent development of digital editing techniques can be used to
synthesize realistic images and videos that could also be used in courts of law.
Unfortunately, photographs taken by amateur photographers are not protected
from tampering. So if these photographs are used as testimony in courts of law,
how is it possible to distinguish true evidence from false one?
In the early days of the Internet, digital watermarking was the main coun-
termeasure against illegal use of digital contents . However, most images and
videos do not have an embedded digital watermark. Once images or videos with-
out watermarks are uploaded to the Internet, digital watermarks are ineffective
even if they are embedded afterwards because the contents may have already
been tampered with by someone. Therefore digital watermarking is found to be
limited in its ability to assure authenticity.
Recently a number of forgery detecting techniques for images without wa-
termarking have been studied . These techniques exploit inconsistencies or
unnaturally high coherence observed in an image. Jonson and Farid used incon-
sistencies in lighting  and chromatic aberration . Lin et al. estimated camera
response function and verified its uniformity across an image . Luk´ aˇ s et al.
extracted fixed pattern noise from an image and compared it with a reference
pattern . Fridrich et al. computed correlation between segments in an image
T. Wada, F. Huang, and S. Lin (Eds.): PSIVT 2009, LNCS 5414, pp. 306–317, 2009.
c ? Springer-Verlag Berlin Heidelberg 2009
Detecting Video Forgeries Based on Noise Characteristics307
and detected cloned regions . Ye et al. used an estimated JPEG quantiza-
tion table and evaluated its consistency . The different digital image forensic
methods mentioned above help us to aggressively estimate the authenticity of
digital images. In contrast, research for digital video forensics is just getting
started, and the development of forgery detecting techniques for video is in high
One of the most frequent digital evidence declared invalid in a court of law is
a video recorded by a fixed surveillance camera. Tampering methods for a scene
that contains a static background can be classified into two approaches. One
is replacing regions or frames with duplicates from the same video sequence:
forgers can hide unfavorable objects in a scene by overwriting these with the
background. The other is clipping objects from other images or video segments
and superimposing them on the desired regions in the video. This type of forgery
aims to show objects that are advantageous for false evidence.
The method for detecting replacement or duplication has been studied by
Wang and Farid . Duplication yields high correlation between original frames
or regions and cloned ones. Detecting unnaturally high coherence is useful for
discovering copy-paste tampering. It has been demonstrated in the research that
we can find substitutions from another frame in the same video sequence. How-
ever, their proposed method has a serious limitation in that it can only detect
copy-paste tampering from the same video sequence. It cannot be used to detect
superimposition, i.e., inserting objects from other video segments. In contrast,
our aim is to propose a method that can detect superimposition.
The basic idea of our proposed method is to use noise inconsistencies between
the original video and superimposed segments to detect forgeries. We exploit the
photon shot noise in a digital camera as a clue to tampering. Photon shot noise
results from the quantum nature of photons and follows a Poisson distribution,
where the variance of the number of photons equals the mean. This dependency
on the irradiance of photon shot noise gives us a clue to inconsistencies in the
video. A CCD camera converts photons into electrons and finally into bits; there-
fore, the relation between the variance and the mean of the number of photons
is converted into that between the variance and the mean of the observed value.
This relation is formulated as the noise level function (NLF) by Liu et al .
The NLF depends on such parameters as inherent parameters of the camera and
recording parameters. Consequently, by comparing the relation of the variance
and the mean in a video clip, we can detect forged regions clipped from another
Specifically, given input video, we first analyze the noise characteristics at each
pixel. Fig.1 shows a diagram of the noise characteristics. The solid line is the
NLF of this distribution. Points in the figure represent the noise characteristic
computed from each pixel. Once we obtain the per-pixel noise characteristics,
NLF is fitted to the points using the least squares method. In this paper, we
assume a linear camera response function (CRF). Since it is known that the linear
CRF yields a linear NLF , the problem of estimating the NLF of the original
video results in the problem of fitting a linear function to the data. We adopt
308M. Kobayashi, T. Okabe, and Y. Sato
Fig.1. Diagram of noise characteristics. Solid line is the estimated noise level function.
Points inside the dashed lines (open circles) are regarded to be authentic. Closed circles
are regarded to be from forged pixels.
the simplest metrics for the forgery measure, i.e., a point whose distance from
the estimated function is greater than a threshold is from a forged pixel. The
dashed lines in Fig.1 are the thresholds that separate the noise characteristic
points into authentic (open circle) pixels and forged (closed circle) pixels. By
evaluated every pixel in this way, we can detect per-pixel forgery in the given
We recorded some real videos for experiments and demonstrated that different
recording parameters resulted in different noise characteristics. Then we applied
the proposed method to the tampered video, we found that our method could
properly detect the forged region.
2 Related Work
2.1 Forgery Detection Methods for Images and Video
The area of digital image forensics has progressed so markedly in the last few
years that several approaches have been developed to detect forgeries in a digital
image. Image tampering methods can be classified into two approaches. One is
replacing regions with others in the same image and the other is superimposing
regions clipped from other images.
The first attempt of forgery detection was proposed by Fridrich et al . This
method targets the copy-move method of attack, which yields unnaturally high
correlation between duplicated regions. The researchers introduced a detection
method based on robust block matching, which was carried out by using Dis-
crete Cosine Transform (DCT) coefficients in order to deal with lossy JPEG
Subsequent approaches target the superimposition-based forgeries, which ver-
ify the uniformity of characteristics in an image; therefore objects clipped from
Detecting Video Forgeries Based on Noise Characteristics309
other images could be detected. Jonson and Farid proposed methods based on
optical clues. They estimated the light source directions from some contours in
an image and checked the consistency of estimated light source directions .
This technique showed so accurate estimation of light source directions for out-
door scenes that it could differentiate tampered objects in the image. Jonson and
Farid also developed a method for detecting forgeries based on lateral chromatic
aberration : a spatial shift of light passing through the optical system due to
the difference of refraction between wavelengths. Global model parameters that
determine the displacement vector at each pixel in an image were estimated,
and the degree of tampering was evaluated by calculating the average angular
error between the displacement vector determined by global parameters and the
actual local vector.
Lin et al. checked for the consistency of the camera response function esti-
mated by analyzing the edges . The irradiance on an edge should be a linear
combination of those from objects at both sides of the edge, but a nonlinear cam-
era response skews the linearity of signal processing. This approach estimates
the nonlinear inverse response functions that convert a nonlinear relation of ob-
served pixel values on the edge into a linear relation. If the function estimated
from an edge does not conform to the rest of the image, the edge is marked as
a sign of tampering.
JPEG is a compression technique for images; different manufacturers design
different quantization tables used in a compression process. Ye et al. proposed
a method to detect inconsistencies in an image based on the blocking artifact
measure . If blocks compressed with different quantization tables are com-
bined in an image, the blocking artifact measure of forged blocks is much larger
than that of an authentic block. They estimated the quantization table from the
histogram of DCT coefficients and evaluated the blocking artifact measure of
Compared to the image forensic techniques mentioned above, only a few tech-
niques have been developed for video. Wang and Farid proposed forgery detect-
ing methods based on video duplication and a deinterlacing algorithm [15,16].
The first approach that detect duplication is similar to the correlation-based de-
tection proposed by Fridrich et al., extended so that it could detect duplicated
regions across frames. They combined spatial and temporal correlation for de-
tecting duplicated frames as well. On the other hand, the deinterlacing algorithm
is a technique of converting interlaced video into a non-interlaced form. Due to
the half resolution of interlaced video, the deinterlacing algorithm makes full
use of insertion, duplication, and interpolation of frames to create full-resolution
video. Parameters in the interpolation and the posterior probability of forgery
are estimated by using the Expectation Maximization (EM) algorithm. Wang
and Farid referred to forgery detection for interlaced videos in the same paper.
They suggested that the motion between fields of a frame is closely related to
that across fields in interlaced videos. Evaluating the interference to this relation
by tampering, they detect the forgeries in the given interlaced video.
310M. Kobayashi, T. Okabe, and Y. Sato
The methods proposed by Wang and Farid are interesting attempts for digital
video forensics. It should be pointed out, however, that these methods have
limitations for forgery detection. The first forensic technique based on correlation
assumes that forged regions are duplicated from the same video sequence. As a
result, this method has the same limitation for forgery detection as the method
proposed by Fridrich et al., that it cannot detect superimposed regions from
other videos. The second method targeting deinterlaced and interlaced videos
can detect superimposing from other video sequences, but it limits the form of
the video to deinterlaced or interlaced form.
Our proposed method is based on the inconsistencies of the noise character-
istics in the given video. Forged regions brought from other video clips can be
effectively detected by our method. In addition, our method exploits the char-
acteristic of camera noise. Noise is a stable clue for forensics because it is an
inevitable phenomenon in signal processing. Therefore our method is applicable
to a wide range of videos.
2.2 Effective Use of Noise in Digital Data
Since the early period of digital camera, various reports have been given on the
study of noise in signal processing. The main purpose of this field of research is
to remove noise in images. Many denoising techniques have been developed and
systematically classified .
On the other hand, some researchers have recently introduced interesting at-
tempts to make effective use of noise, rather than trying to remove it from images
and videos. Matsushita and Lin exploited the distribution of noise intensity for
each scene irradiance to estimate the camera response functions (CRFs) .
Noise distribution is by nature shown to be symmetric, but it is skewed by non-
linear CRFs. Conversely, the inverse CRF can be estimated by evaluating the
degree of symmetry of back-projected irradiance distribution. Using the noise in
an image, the detection ability of the method is not degraded by noise and thus
the method can be used under conditions of high-level noise.
Liu et al. estimated the noise level function (NLF) from a single image, which
relates the noise intensity with the image intensity . The spatial variance
in an image contains the variance resulted in object’s texture as well as the
intensity of the noise. Obtaining the component of the real noise from NLF, we
can disassociate the component of texture from the variance of the observation.
They utilized the function not only for denoising but also for adaptive bilateral
filtering and edge detection.
Noise information is available for camera identification and forgery detection
as well. Due to the sensor imperfections developed in a manufacturing process,
the CCD camera contains pixels with different sensitivity to light. This spatial
variation of sensitivity is temporally fixed and known as fixed pattern noise. Since
this non-uniformity is inherent in a camera, we can exploit it as a fingerprint.
Luk´ aˇ s et al. determined the reference noise pattern of a camera by averaging the
noise extracted from several images . They extracted fixed pattern noise from
a given image using a smoothing filter and identified the camera that took the
Detecting Video Forgeries Based on Noise Characteristics311
image. The authors also proposed a method for detecting forgeries in an image
using the same approach .
This paper introduces a video forensic method by checking for inconsistency
of the noise characteristics, which has never been proposed among the forensic
methods for videos. Since the proposed method aggressively exploits noise, it is
effective also for a video contaminated by significant noise. Other approaches are
not able to handle high levels of noise.
In this section, we propose a forgery detecting method using a noise character-
istics model. In this paper, we will consider the inconsistencies of the charac-
teristics of the noise mixed in the signal to be a clue to tampering. We first
introduce a noise characteristic model in Section 3.1. As stated before, we fo-
cus in particular on photon shot noise for detecting forgeries in the given video.
This is because the variance of observed intensity caused by photon shot noise is
closely related to its mean. The relationship between the variance and mean of
observed intensity is formulated as the noise level function (NLF), which is the
clue to tampering. In Section 3.2, we propose a method to estimate NLF and
detect forgeries by using the estimated NLF.
3.1Noise Level Function of Video
A CCD digital camera converts photons into electrons and finally into bits. This
signal processing has been studied for a long time [3,13]. In the signal process
of a digital camera, several noise sources corrupt the signal such as photon shot
noise, dark current noise, thermal noise, read-out noise and quantization noise.
We focus on photon shot noise among these noise sources because of the following
two reasons: (1) photon shot noise is dominant noise in a scene except in an
extremely dark environment, and (2) the relation between the brightness and
the noise intensity is useful for forgery detection.
The number of photons that enters a CCD element has temporal fluctuation
and thus this variation behaves as noise. Since this fluctuation follows a Poisson
distribution, the noise intensity depends on its mean – the noiseless irradiance.
Unfortunately, we cannot measure the distribution of photons directly because
photons are converted into electrons, electric voltage, and finally bit chains.
However, we can instead compute the relation between the mean and the variance
of the observed pixel value. We consider their relation as a measure of tampering.
LetˆO be the noiseless observed intensity. Due to the effect of noise, the real
observation has fluctuation and thus we obtain a random variable of observation
O. Let μˆ Oand σ2
O, respectively, when the noiseless observation isˆO. Following the formulation
described in , we introduce NLF τ(μˆ O) as
ˆ Obe the mean and the variance of the observed pixel intensity
τ(μˆ O) = E[(O − μˆ O)2].
312 M. Kobayashi, T. Okabe, and Y. Sato
Unlike the equation in , we do NOT calculate the square root of Mean Square
Error. This function represents how the variance changes with respect to the
mean of the observed pixel value. When we obtain the mean observation μˆ O, the
variance is described by a function with respect to the mean as
NLF depends on such parameters as inherent parameters of the camera and
recording circumstance; they include inherent parameters of a camera such as
quantum efficiency and the response function, and recording parameters such as
exposure and electric gain.
For the sake of simplicity, we make two assumptions regarding the input video.
The first assumption is that the distribution of the noise is zero-mean, and
therefore we can obtain noiseless observed intensity of each pixel by averaging.
Since this assumption suggests that the mean of observed intensity equals the
noiseless intensity, we rewrite μˆ Oas simply μ. Second, we assume a linear camera
response function (CRF). Former research on noise in a CCD camera  implies
that a linear CRF yields a linear NLF. Therefore we simply apply linear least
squares method to the calculated points.
ˆ O= τ(μˆ O).
Based on the theoretical background described in the previous section, we an-
alyze the noise characteristics and detect forgeries of the given video by the
following process. First, the mean and the variance of the pixel value are calcu-
lated at each pixel. Next, the NLF is estimated by fitting a function to the noise
characteristic points. Finally, each pixel is evaluated based on its distance from
the estimated NLF. We describe each step in detail in the following.
Detection of Forged Pixels
Calculation of noise characteristics.
frame of video sequence, NLF can be obtained by calculating spatial mean and
variance. This approach, however, requires an assumption of the local uniformity
of the object’s reflectance and shading. If there is a textured object in a scene,
we cannot obtain the noise component independently from the total variance
because the spatial variation is mixed in the signal. The proposed method proves
its merits in this case. As mentioned in the introduction, we deal with a static
scene where the camera and the objects are fixed during recording. Therefore
a conclusion is drawn that the temporal variation of each pixel value results
entirely from noise. Operating statistical analysis along a time-line to the given
video, we obtain the relation between μ and σ2
If we have an image or a single
ˆ Oat each pixel.
a dense set of points, as many as the resolution of the video. Then we fit a linear
NLF τ(μ) to the points using linear least squares method as
τ(μ) = αμ + β,
where α and β are the estimated parameters. In order to eliminate the effect of
the scale factor between the mean and the variance, they are normalized before
Analyzing observed intensity along a time-line, we obtain
Detecting Video Forgeries Based on Noise Characteristics313
Fig.2. Example of the recorded video
Table 1. Recording parameters
Fig.3. Noise characteristics with different
gain. Data points are thinned out for dis-
play. Shutter times and gain of data sets
are shown in Table 1.
Because the noise intensity of the video created from an authentic process is
uniquely determined by the estimated NLF, every pixel value converted from
the same irradiance should yield the same noise intensity. Consequently, incon-
sistencies of the relation between the mean and the variance can be a clue to
the forgery. Therefore we can claim pixels whose noise characteristic is far from
NLF to be from a tampering process.
In this paper we use RANSAC  so that the NLF is estimated robust to the
outliers calculated from the forged regions. The closed circles in Fig.1 are the
outliers. Although we need to set a threshold manually, RANSAC is relatively
robust to outliers, considering its ease of implementation.
Evaluation of pixels.
each pixel in the video is evaluated based on the distance from the estimated
NLF according to (2). The evaluation of the pixel N located at the position r
is determined as follows.
Once we obtain the NLF τ(μ), the authenticity of
ˆ O(r) − τ(μ(r))
??? > ε
where ε is the constant threshold.
Note that near the maximum pixel value (Here we consider 8-bit depth, hence
the maximum is 255), the observed values are saturated and their apparent
variances are smaller than real ones, which causes degradation of the detection
quality. Therefore we set an upper limit T for the mean value to omit evaluation
of the pixels with the mean larger than T.
314 M. Kobayashi, T. Okabe, and Y. Sato
4 Experimental Results
All the experiments were done on video recorded on a PointGreyFlea digital cam-
era. 128 grayscale frames are recorded at 30 fps for the 640 × 480-resolution com-
Boardunder sunlight as the object. Fig. 2shows an exampleof the recordedvideo.
4.1Noise Characteristics with Various Parameters
We first showed how the noise characteristics change based on the recording
parameters. Fig.3 shows the comparison of the noise characteristics with various
electric gain. The shutter times and the gain of the data sets are shown in Table
1. Note that the horizontal and vertical axes indicate absolute, not normalized,
values. The data points of each set distribute on a line that rises steeply corre-
sponding to the gain. In the range of upper limit, the variances fall rapidly to
zero, which results from the saturation in the quantization process.
4.2Forgery Detection Using Noise Characteristics
We conducted another experiment of forgery detection. We created forged video
clips as follows from 6 video sources analyzed above. At first a pair of videos
taken under different parameters was chosen from the sources: they are a pair of
the original and the replaced video clips. A forged region of 100×100-dimension
was randomly located, and the position was kept as the ground truth. The pixel
values in the located region over all frames in the original video were overwritten
by those in the replaced video. An example of a frame in the forged video is shown
in Fig.5 (Left). The white box in the image indicates the forged region.
The noise characteristics of the forged video were calculated as described in
the previous section. Fig.4(A) shows the noise characteristics of the forged video
created by replacing a part of the video of parameter (a) with that of parameter
(c) in Table 1. Note that the means and the variances are normalized in this
figure. Using RANSAC, we fitted a linear NLF to the calculated points. The
threshold parameter of RANSAC was empirically set to 0.1 in the normalized
noise characteristics space. There are two clusters: a dense cluster projected from
the region of parameter (a) and a sparse cluster from the region of parameter
(c). The solid line in the figure is the estimated NLF. Due to RANSAC, the
linear NLF is properly estimated robust to the outliers.
of forgery ε is set to 0.1, which is equal to the threshold of inliers on RANSAC.
The upper limit of the mean value for evaluation T is empirically set to 0.9.
Fig.5 (Right) shows the detection result for the test data shown in Fig.5
(Left). The highlighted pixels in the figure represent the pixels determined to
be forged. The proposed method detects most of the forged pixels in the color
patches, while some pixels in the border are accepted. This is because the noise
characteristic in the dark border is not sufficiently distinctive from that of the
pixels in the authentic region to differentiate between them.
Detecting Video Forgeries Based on Noise Characteristics315
Fig.4. Noise characteristics of a mixture video containing parameter (a) for the original
and (c) for the replaced region (A) and vice versa (B). The solid line is the estimated
NLF by using RANSAC and the dashed lines are the boundaries of forgery.
Fig.5. Left: Example of the forged video. White box indicates the forged region. Right:
Detection result for the video shown in the left figure. Highlighted pixels are determined
to be forged.
To evaluate our method, we calculated the recall and the precision rates for
every combination of the video clips. For one set of the recording parameters, we
averaged over 30 random trials. The parameters in the fitting and the detection
process were constant over this evaluation. The experimental result is shown in
Table 2. We found that the proposed method can differentiate the forged pixels
when the noise characteristics in the forged region are sufficiently isolated from
the rest of the video.
However, the proposed method does not evaluate the authenticity of the pixels
brighter than the upper limit T, which may cause degradation of detection. Even
in the case that the noise characteristics are well separated, recall becomes worse
if the forged region is located on a bright color patch. In addition, we should take
notice of the low precision rate in the lower triangular portion of the table. The
316M. Kobayashi, T. Okabe, and Y. Sato
Table 2. Evaluation result (Top: Recall [%], Bottom: Precision [%])
characteristicpoints of the original video in these conditions spread broad in spite
of the constant boundary of forgery (See Fig.4(B) for an example). That is why
there occurred many false-positives and the quality of the detection is degraded.
It should be noted that the threshold for outliers in RANSAC is empirically ad-
justed and constantwith the gain.Nevertheless,the proposedmethod achievesro-
bust fitting for all combinationsof the recordingparametersbecause of the benefit
of robust fitting. It is interesting that the parameters can be fixed because we can
easilydetect forgeriesproperlywithout aprobabilisticmodel oradaptivelearning.
5Conclusions and Future Work
In this paper we introduce a noise level function of a video clip and propose a
digital video forensic technique based on the noise characteristics. The proposed
method calculates the noise characteristic of each pixel by using temporal aver-
aging, and achieves per-pixel evaluation of the authenticity with a high degree
of accuracy by using a fitting method robust to outliers.
The following considerations will provide work for the future. First, in this
paper we deal only with the videos recorded from a static scene, but in the future
we will definitely have to consider working with persons and moving objects. In
addition, the spatial relation of pixels is not used in this paper, but it will be
useful for locating objects to integrate information of neighboring pixels. Also,
combined with image segmentation techniques, it is expected that the method
will reveal suspicious regions in the given video. Second, nonlinear CRFs are not
considered in this report. In order to apply our method to a variety of cameras,
we should expand it to generalized NLFs.
Detecting Video Forgeries Based on Noise Characteristics 317
1. Fischler, M.A., Bolles, R.C.: Random sample consensus: a paradigm for model
fitting with applications to image analysis and automated cartography 24(6), 381–
2. Fridrich, J., Soukal, D., Luk´ aˇ s, J.: Detection of copy-move forgery in digital images.
In: Proc. of Digital Forensic Research Workshop (2003)
3. Healey, G.E., Kondepudy, R.: Radiometric ccd camera calibration and noise esti-
mation. IEEE Transactions on Pattern Analysis and Machine Intelligence 16(3),
4. Johnson, M.K., Farid, H.: Exposing digital forgeries by detecting inconsistencies
in lighting. In: Proc. of Workshop on Multimedia and security (2005)
5. Johnson, M.K., Farid, H.: Exposing digital forgeries through chromatic aberration.
In: Proc. of International Multimedia Conference, pp. 48–55 (2006)
6. Lee, S.-J., Jung, S.-H.: A survey of watermarking techniques applied to multimedia.
In: Proc. of IEEE International Symposium on Industrial Electronics, vol. 1, pp.
7. Lin, Z., Wang, R., Tang, X., Shum, H.-Y.: Detecting doctored images using camera
response normality and consistency. In: Proc. of IEEE Computer Society Confer-
ence on Computer Vision and Pattern Recognition, vol. 1, pp. 1087–1092 (2005)
8. Liu, C., Szeliski, R., Kang, S.B., Lawrence Zitnick, C., Freeman, W.T.: Automatic
estimation and removal of noise from a single image. Technical Report MSR-TR-
2006-180, Microsoft Research (December 2006)
9. Luk´ aˇ s, J., Fridrich, J., Goljan, M.: Determining digital image origin using sensor
imperfections. In: Proc. of Society of Photo-Optical lnstrumentation Engineers
Conference, vol. 5685, pp. 249–260 (2005)
10. Luk´ aˇ s, J., Fridrich, J., Goljan, M.: Detecting digital image forgeries using sensor
pattern noise. In: Proc. of Society of Photo-Optical Instrumentation Engineers
Conference, vol. 6072, pp. 362–372 (2006)
11. Matsushita, Y., Lin, S.: Radiometric calibration from noise distributions. In: Proc.
of IEEE Computer Society Conference on Computer Vision and Pattern Recogni-
tion, pp. 1–8 (2007)
12. Motwani, M.C., Gadiya, M.C., Motwani, R.C., Harris Jr., F.C.: Survey of image
denoising techniques. In: Proc. of Global Signal Processing Expo. and Conference
13. Tsin, Y., Ramesh, V., Kanade, T.: Statistical calibration of ccd imaging process. In:
Proc. of IEEE International Conference on Computer Vision, vol. 1, pp. 480–487
14. Van Lanh, T., Chong, K.-S., Emmanuel, S., Kankanhalli, M.S.: A survey on digital
camera image forensic methods. In: Proc. of IEEE International Conference on
Multimedia and Expo., pp. 16–19 (2007)
15. Wang, W., Farid, H.: Exposing digital forgeries in interlaced and deinterlaced video.
IEEE Transactions on Information Forensics and Security 2(3), 438–449 (2007)
16. Wang, W., Farid, H.: Exposing digital forgeries in video by detecting duplication.
In: Proc. of Workshop on Multimedia & security in International Multimedia Con-
ference, pp. 35–42 (2007)
17. Ye, S., Sun, Q., Chang, E.-C.: Detecting digital image forgeries by measuring in-
consistencies of blocking artifact. In: Proc. of IEEE International Conference on
Multimedia and Expo., pp. 12–15 (2007)