Examining the Performance for Forensic Detection of
Rare Videos Under Time Constraints
Johan Garcia
Department of Mathematics and Computer Science, Karlstad University, Karlstad, Sweden
Keywords:
Digital Forensics, Video Classification, Monte-Carlo Simulations.
Abstract:
In many digital forensic investigations large amounts of material needs to be examined. Investigations in-
volving video files are one instance where the amounts of material can be very large. To aid in examinations
involving video, automated tools for video content classification can be employed. In this work we examine
the performance of several different video classifiers in the context of forensic detection of a small number
of relevant videos among a large number of irrelevant videos. The higher level task performance that is of
interest is thus the ability to detect a relevant video in a limited amount of time. The performance on this
higher level task is a combination of the classification performance, but also the run-time performance of the
classifiers. A variety of video classification techniques are available in the literature. This work examines
task performance for 6 video classification approaches from literature using Monte-Carlo simulations. The
results illustrate the interdependence between run-time and classification performance, and show that high
classification performance in terms of true positive and false positive rates not necessarily lead to high task
performance.
1 INTRODUCTION
Video content classification has been a topic of in-
terest to the research community for a considerable
amount of time. Many different approaches exist
which differs in classification and runtime perfor-
mance. One area where video content classification
may be beneficial is in digital forensics. When per-
forming forensics examinations, the amounts of data
that needs to be handled are often substantial. By em-
ploying various automated classification tools, exam-
inations can be performed more efficiently.
In this work we focus on examinations that are
performed under time constraints. One instance when
this could happen is when there is a suspicion of inap-
propriate material, such as adult videos, on computers
or storage devices belonging to a school or an organi-
zation where such material is not allowed on the orga-
nizations equipment. Another case where time con-
straints may be present is during the initial phase of
investigationsrelated to child sexual abuse (CSA) ma-
terial. Such examinations can take place when there
is a degree of suspicion but not sufficiently to warrant
the seizure of equipment. The equipment might be ex-
amined in place for a period of time and if data of rel-
evance is located then this would be grounds for col-
lecting the equipment and making a thorough forensic
analysis.
The particular task that the video classification
should support in these contexts is the detection of
a small number of relevant videos among a large
number of non-relevant videos. This particular use
case presents particular challenges to the video clas-
sifier with regards to classification and runtime per-
formance. In many cases video classifiers were de-
signed and evaluated on classification performance
and with no, or little, consideration of runtime charac-
teristics. However, since the runtime performance of
video classification approaches can vary several or-
ders of magnitude it can be of considerable impor-
tance for task performance in a time constrained set-
ting. In this paper we examine six different video
classification approaches in the context of detecting
rare relevant videos that occur with low frequency
among a large number of irrelevant videos. In this ex-
amination the classifiers consider the relevant videos
to be detected adult videos among a set of regular
videos. However, the same trade-off between classi-
fication and runtime performance occurs for the case
where a small number of CSA videos should be de-
tected among a large number of non-CSA videos. To
perform the evaluation of the video detection perfor-
419
Garcia J..
Examining the Performance for Forensic Detection of Rare Videos Under Time Constraints.
DOI: 10.5220/0005574204190426
In Proceedings of the 12th International Conference on Security and Cryptography (SECRYPT-2015), pages 419-426
ISBN: 978-989-758-117-5
Copyright
c
2015 SCITEPRESS (Science and Technology Publications, Lda.)
mance Monte Carlo simulations are employed. The
results show that for the considered classification ap-
proaches the difference in runtime characteristics are
so large that they dominate over the classification per-
formance in the majority of settings. Only when the
number of rare videos that can be detected is very
low, a slower but better-classifying approach is more
appropriate for a subset of the considered time con-
straint times.
The paper is structured as follows. In the next sec-
tion the background and related works are discussed.
Section 3 covers the evaluation setup, followed by the
results in Section 4. Finally, Section 5 provides the
conclusions.
2 BACKGROUND
A problem when comparing different video classifi-
cation approaches is that the papers describing the
different approaches use different data sets. A valid
comparison of the classification and run-time perfor-
mance of the different approaches can not be achieved
when different data sets and different hardware are
employed by the different authors. As the video de-
tection task is dependent on data for both the video
classification performance and the runtime perfor-
mance under comparableconditions the amount of us-
able input from the literature is limited. In some cases
authors also re-implement approaches described ear-
lier in the literature and evaluate this collection of ap-
proaches on the same data sets and hardware. This is
the case for the work described in (Jung et al., 2014).
Our present study is based on the classification and
performance results reported by them. In their paper
Jung et al. consider eight different video classifica-
tion approaches involving various characteristics of
the videos with some focus on the presence of peri-
odic motion. Out of these eight, two are proposed by
the authors.
The Jung1 approach proposed by the authors
are based on classification performed by a 70-
dimensional support vector machine (SVM). The fea-
tures used are based on the magnitude and frequency
of periodic motion along with a skin color feature
vector based on hue and saturation histograms. They
also propose the Jung2 approach which is similar but
does not consider the skin color information. In ad-
dition to their approach, Jung et al. have also im-
plemented an approach suggested in (Behrad et al.,
2012). The Behrad approach is based on locating the
largest section that is a skin colored, and compute the
correlation between frames, and subsequently analyze
the discrete Fourier transform of the correlation. An-
other examined approach suggested in (Ulges et al.,
2012) is based on fusing several features including
skin color, the discrete cosine transform coefficients
of image patches using a bag of visual words model,
motion histograms and also audio information based
on cepstral coefficients. This multitude of employed
features comes with a corresponding impact on clas-
sification speed. Another approach is suggested in
(Endeshaw et al., 2008) which has low computational
requirements as it works only on the motion vectors
present in the compressed video stream. While such
approach has considerable runtime benefits, it was
also found to yield worse classification performance
than the previously discussed approaches. In addition
to these approaches, Jung et al. also evaluates three
schemes designed to detect periodic motions. These
approaches (Briassouli and Ahuja, 2007; Niyogi and
Adelson, 1994; Liu and Picard, 1998) were not orig-
inally designed with any consideration of using them
for classifying adult video material.
While the current study focuses on the detection
of adult videos, the same consideration regarding the
trade-off between runtime performance and classifi-
cation performance exists in the domain of CSA in-
vestigations. As long as the rare video detection
task can be achieved using different approaches with
varying classification and runtime characteristics the
same methodology as used in this study is applica-
ble. The paper (de Castro Polastro and da Silva Eleu-
terio, 2012) highlights the importance of being able
to quickly detect videos in the context of CSA in-
vestigations. The automated classification of CSA
video material is discussed in (Schulze et al., 2014),
which examines a multi-modal approach. Hence the
possibility exists to examine the resulting task per-
formance for the different modalities based on their
varying run-time and classification characteristics in
a manner similar to what is done here.
With regards to Monte Carlo simulations as em-
ployed here, one example where it has been used pre-
viously in the domain of digital forensics is (Garcia,
2014). There, the implications of including side in-
formation for file hashing has been explored using ap-
proaches similar to what is used in this study but with
more focus on storage device characteristics.
3 EVALUATION SETUP
3.1 Performance of Classifiers
The video classification performance for the eight
classifiers discussed in the background section are
shown in Figure 1. The figure shows the Receiver Op-
SECRYPT2015-InternationalConferenceonSecurityandCryptography
420
Figure 1: Classification performance (from (Jung et al.,
2014)).
erating Curve (ROC) which presents a curve of all the
possible operating points of a binary classifier. The x-
axis shows the false alarm rate, or false positive rate
(FP), which in this case corresponds to the fraction of
videos classified as being relevant although they actu-
ally are irrelevant. The y-axis shows the true positive
rate (TP), which is the fraction of the relevant videos
that are actually detected as such. Ideally, FP should
be low and TP high, leading to the upper left corner
representing the optimal classifier. The curves in the
graphs represents the possible trade-offs between FP
and TP that results from setting the detection thresh-
old for an individual classifier to different values. By
setting the threshold to a specific value the classifier
will work at a particular operating point with specific
FP and TP.
In addition to the classification performance
shown above, runtime aspects are also of importance
in this evaluation. The run time characteristics of the
eight classifiers as reported in (Jung et al., 2014) are
shown in Table 1. The table shows the amount of
Table 1: Run time performance (times are from (Jung et al.,
2014)).
Method Processing time Relative speed
for 6 hours video
Jung1 9 min 30 sec 37.90x
Ulges 3 h 17 min 1.83x
Behrad 18 min 20 sec 19.64x
Jung2 9 min 9 sec 39.34x
Endeshaw 1 min 21 sec 266.67x
Briassouli 5 h 33 min 1.08x
Niyogi 2 h 32 min 2.37x
Liu 2 h 26 min 2.47x
time necessary for processing video files with the to-
tal duration of six hours, and the resulting speed of
the classifier in relation to normal real-time. These
results were obtained by Jung et al. using a PC with
a 2.8GHz CPU and 4GB of RAM. The videos had a
resolution of 640× 480 pixels. As can be seen in the
table, there is a large variation between the different
classification approaches. These differences relate to
how the classification is performed, as discussed in
the background section.
3.2 Chosen Parameter Setup
Based on their performance characteristics, six classi-
fication approaches were selected for a further Monte
Carlo simulation-based performance evaluation. As
the Niyogi and Liu approaches had relatively poor
classification performance for this task, and had no
particular runtime advantages they were not con-
sidered in the further evaluation. Based on the
ROC classification performance characteristics pro-
vided above,suitable operatingpoints for the different
remaining classification approaches were decided us-
ing heuristics. For this particular task operating with
a low false positive rate is considered more impor-
tant than having the highest possible true positive rate.
The chosen operating points are shown in Table 2.
Table 2: Selected operating points.
Method True positive False positive
rate rate
Jung1 0.90 0.013
Ulges 0.83 0.013
Behrad 0.76 0.008
Jung2 0.70 0.015
Endeshaw 0.63 0.015
Briassouli 0.60 0.070
In addition to classifier characteristics, the sim-
ulator also needs to model the characteristics of the
video material to be evaluated. This study focuses on
the case where a small number of relevant videos is
to be detected among a large number of non-relevant
videos. The videos are modeled as coming from
three different classes with different characteristics.
The first class represents material obtained via video
sharing sites such as YouTube. The YouTube statis-
tics is also considered to represent amateur-generated
content in general. Statistics for the run time, com-
pression rate, and byte length of videos present on
YouTube are reported in (Ameigeiras et al., 2012).
The T3 data set provided by that study is also used
here, although it was filtered to only consider videos
with a width of 480 pixels. The two other classes con-
sidered are downloaded TV shows and movies. For
these classes the byte length distribution reported in
(Carlsson et al., 2012) is used in a simplified way, us-
ing the reported medians, but with a uniform distribu-
ExaminingthePerformanceforForensicDetectionofRareVideosUnderTimeConstraints
421
Figure 2: Video file size distribution.
Figure 3: Compression distribution.
tion of ±30% for TV shows and ±40% for movies.
As no empirical data was found for the compres-
sion rate of these classes, it was set to vary uni-
formly within the range 765-1035 kbps. The rel-
ative proportion of video files were set to 0.4 for
YouTube/amateur videos, 0.3 for TV shows, and 0.3
for movies.
For these proportions, the resulting cumulative
density function (CDF) for the variation of byte length
file sizes in the modeled video material is shown in
Figure 2. The leftmost hump in the figure corre-
sponds to the YouTube/amateur material, and the two
rightmost slopes correspond to TV shows and movies.
Here, the x-axis is logarithmic. Similarly, the result-
ing compression rate CDF is shown in Figure 3. Fi-
nally, the modeled video duration CDF is shown in
Figure 4. Here, the part under 1000 seconds comes
mainly from the empirical time distribution found in
the YouTube data set, whereas the part above 1000
seconds mainly model the durations of the TV series
and movie classes. As can be seen in the figure the
Figure 4: Video duration distribution.
median time for modeled TV show is around 30 min.,
and for the movie class around 1 hour 40 min.
3.3 Simulation Configuration
The simulation configuration models 500 GB of video
material stored on a storage device. As the bottleneck
in this study is classifier runtime performance rather
than storage device read characteristics, no storage
device related delays were included in the simula-
tion model. One Monte-Carlo simulation round con-
sists of filling the space on the storage device with
video files with the characteristics given by the file
size, compression rate, and duration distributions dis-
cussed in the previous subsection. Then, a particu-
lar number or fraction of files are randomly marked
as relevant, or rare. The different evaluation metrics
are then computed using the aggregated values after
considering the operating point and run-time charac-
teristics for each of the classifiers for each file. The
simulations used 10000 Monte-Carlo runs. Each sim-
ulation run generates an average of around 1800 video
files.
3.4 Validity Aspects
Although data from literature was employed for sev-
eral parameters and model aspects for the simula-
tion, it should be recognized that some parametriza-
tion simplifications were made. Some simplifications
were performed by Jung et al when implementing
some of the classification approaches as discussed in
their paper.
A sensitivity analysis using different video dis-
tributions and operating points showed that the rela-
tive performance of the examined approaches is ro-
bust to such variations in the parameterization. Thus,
these results are considered to be sufficiently general
to strongly support that run-time characteristics needs
more attention when designing classifiers intended for
the forensic domain.
4 RESULTS
The results section is divided into three parts. In the
first part the focus is on the amount of time until the
first detection of a rare video. The performance here
depends mainly on runtime characteristics, but when
the number of rare videos is very small the true pos-
itive rate is also of importance. The second part con-
siders the evolution of the number of detected videos
overtime, and the amount of manualverification work
SECRYPT2015-InternationalConferenceonSecurityandCryptography
422
Figure 5: CDF of time to first detection of a rare video, with a 0.01 fraction of rare videos.
that is necessary. The third part shows how a com-
bination of classifiers can be employed to create im-
proved performance for the metrics discussed in the
first two parts.
4.1 Time to First Detect
The time required to detect the first rare video de-
pends on where in the sequence of examined videos
the first rare video happens to be placed. In reality
this varies from case to case, and this variation is cap-
tured by the Monte-Carlo simulations which random-
izes the placement between each simulation run. The
first examination considers how the time to detect the
first rare video varies as a result of the randomness
in the placement of the rare files and other random
factors. The results for the different classifiers when
the fraction of rare files is 0.01 are shown as a CDF
in Figure 5. Note the logarithmic x-axis. As can be
seen in the figure, the amount of clock time to detect
the first video varies considerably. The median time
varies from ca 20 min to ca 83 hours. In a practical
setting such difference can have a considerable im-
pact.
The CDF can also be used to infer the probability
of detecting one rare video in a particular time. If a
time constraint of 1 hour (=3.6× 10
3
s) is present, the
probability of detection varies between 92% and 1%.
This difference is mainly dependent on the run-time
characteristics of the classifier, rather than its classifi-
cation performance. As one focus with this examina-
tion is how video classification approaches function
under time constraints, this graph can give an indi-
Figure 6: CDF of time to detection of a single rare video,
when only one single rare video is present.
cation of how appropriate different classification ap-
proaches would be for particular time limits.
Figure 5 shows the conditions for a rare video
fraction of 0.01. If this value is changed, this af-
fects the time to detect the first video. As one ex-
treme point, it is possible to examine the characteris-
tics when there is only one rare video stored among
the large (ca 1800) number of non-relevant videos.
The case of a single rare video is shown in Figure 6,
which shows the corresponding CDF. Here it is no-
ticeable that the detection times have moved consid-
erably to the left in the figure, making the average
time it takes to detect the video considerably longer.
Also visible in the figure is the fact that for a single
rare video the true positive rate of the classification
approach has a marked impact on the results. As can
be seen in the figure the approaches tops out at a CDF
probability that is equal to their true positive rate. For
the fastest Endeshaw approach this means that among
ExaminingthePerformanceforForensicDetectionofRareVideosUnderTimeConstraints
423
Figure 7: Probability to detect at least one rare video.
the 10,000 simulation runs, for a fraction of 0.37 of
these runs the single rare file was not detected by the
classifier. While this is a weakness of the Endeshaw
classifier, it can nevertheless be observed that for a
large fraction of the considered time periods it is the
best choice. For all time periods below 8.5 × 10
4
s
(≈ 24 hours) the Endeshaw classifier has the highest
probability of detecting the video.
In most cases it is reasonable to expect that the
number of rare videos to be detected is larger than
one. Based on the true positive rates it is possible
to calculate the probability of detecting at least one
rare video as a function of the number of rare videos
stored. The results are illustrated in Figure 7, which
shows that when there are 4-5 rare videos present, the
difference in probability of detecting at least one rare
video becomes very small.
4.2 Detections Over Time
It is possible to examine how the number of detected
videos vary as a function of time for the different
classifiers. Such data makes it possible to consider
the performance implications of using another thresh-
old than one video. The median number of detected
videos over time is shown in Figure 8. The time to
detect the first video in this figure corresponds to the
median value of the CDF in Figure 5. While the Ende-
shaw classifier will be able to detect the most videos
for a large fraction of the time period, it cannot detect
as many videos as other classifiers due to its lower
TP rate. In addition to examining the median, it is
also possible to consider the 95th percentile of the
10000 MC runs. This is shown in Figure 9 which
shows that both the time to detection as well as the
maximum number of detected videos is lower, as can
be expected. It can be noted that the transition point
where Jung1 performs better than Endeshaw is set to
similar point in time for both the median and 95th per-
centile.
Another factor to be considered is the amount of
time that has to be spent validating the videos classi-
Figure 8: Number of detected videos over time (50th per-
centile).
Figure 9: Number of detected videos over time (95th per-
centile).
fied as relevant. As the false positive rate at the cho-
sen operating points vary, the number of false posi-
tive videos also vary. The relative classification speed
also plays a role, as the faster the classification is per-
formed, the more videos can potentially be misclas-
sified. The resulting validation amount is shown in
Table 3. After one clock hour the Jung1 approach has
Table 3: Generated validation amounts.
Method FP Relative Validation
rate speed amount
Endeshaw 0.015 266.67x 4x
Jung1 0.013 37.89x 0.493x
Jung2 0.015 39.34x 0.590x
Behrad 0.008 19.64x 0.157x
Ulges 0.013 1.83x 0.024x
Briassouli 0.070 1.08x 0.076x
scanned circa 38 hours of video and generated an av-
erage of 29 minutes of video which needs to be manu-
ally verified, in addition to any true positivesdetected.
The faster and less precise Endeshaw approach has
during one hour scanned around 267 hours of video
SECRYPT2015-InternationalConferenceonSecurityandCryptography
424
Figure 10: Endeshaw-Jung1 cascading classifier approach.
and generated 4 hours of video material that needs
verification. Out of all detected material some are ac-
tual true positives containing relevant material, and
others are irrelevant material that the classifier mis-
classified. To separate these true positives from the
false positives manual validation is necessary. The
manual validation can be performed in different ways
as deemed appropriate, but viewing the video with a
moderate amount of fast forward is one simple ap-
proach.
4.3 Cascading Classifiers
In addition to using a single classifier it is also pos-
sible to cascade several classifiers after each other.
This is done to improve some aspect of the classifica-
tion system, although at the expense of having a more
complex system. An early example of using cascad-
ing classifiers is the face detection approach described
by (Viola and Jones, 2001). This subsection examines
the performance effects of using a cascading approach
for rare video detection, with the results derived un-
der the assumption of classifier independence. The
results in this subsection are thus indicative of a best
case cascading scenario. As the modes of operation
of the cascaded classifiers are significantly different
they are expected to be quite independent. The exam-
ined cascading approach is shown in Figure 10. The
fast Endeshaw classifier is first employed, and for ev-
ery video classified as positive the Jung1 classifier is
then directly employed as shown in the upper part of
Figure 10. This cascading reduces the amount of false
positives and the corresponding need for manual ver-
ification. Thus, the amount of video per clock hour
needing manual verification is reduced from 4 hours
to less than 10 minutes. The downside is a slight re-
duction in relative speed, which still is above 200x.
Furthermore, the approach of Figure 10 employs
a second run after the first run of all videos has fin-
ished. This second run uses the Jung1 classifier with
the goal of detecting the rare videos among the nega-
tives returned by the Endeshaw classifier. This way a
Figure 11: CDF of time to detect a single rare video, includ-
ing an Endeshaw-Jung1 cascading classifier.
Figure 12: Median number of detected videos over time,
including an Endeshaw-Jung1 cascading classifier.
larger fraction of the total number of rare videos can
be detected, and it will also increase the probability of
finding a single rare video to the same level as for the
Jung1 classifier.
The resulting performance of the cascading ap-
proach for the case of one single rare video is shown
in Figure 11. In the figure the increased probability of
detecting the video is evident, as well as a slight time
increase in relation to the Endeshaw approach. The
knee visible for the cascading classifier corresponds
to the changeover point between videos classified as
positive in the first or second run.
Figure 12 shows the median number of detected
videos for a 0.01 fraction of rare videos. The results
show a similar trend as was shown in Figure 11, with
a slight increase in time and a higher number of de-
tected videos. If the decrease in manual validation
amount provided by the upper classifier of Figure 10
is not required, the upper classifier cascade can be re-
moved. Such removal would lead to run-times below
the knee being similar to those of the Endeshaw clas-
sifier, while still being able to detect the same amount
of videos as the Jung1 approach.
ExaminingthePerformanceforForensicDetectionofRareVideosUnderTimeConstraints
425
5 CONCLUSIONS
Various classification approaches are used in a vari-
ety of contexts. Some of these contexts involve time-
constraints where the run-time characteristics of the
classification approach becomes especially relevant.
In this work we have studied the higher level task
of detecting a rare video with relevant characteristics
among a large set of irrelevant videos. For such tasks
the run time characteristics can be expected to have
high importance in addition to the classification per-
formance.
In this study Monte-Carlo simulations were per-
formed to evaluate the task completion time as re-
lated to the run-time and classification performances
of six different video classification approaches. Per-
formance using cascading classifiers were also exam-
ined. The presented study reflects only a subset of the
available approaches and has some simplifications. It
nevertheless provides strong support for the statement
that classification performance cannot be the only fac-
tor in considerationwhen designing classification sys-
tems used in time constrained environments. The
results show that high classification performance in
terms of true positive and false positive rates not nec-
essarily lead to high task performance.
ACKNOWLEDGEMENTS
The authors wish to thank Soonhung Jung for sharing
data on the receiver operating conditions from their
study.
REFERENCES
Ameigeiras, P., Ramos-Munoz, J. J., Navarro-Ortiz, J., and
Lopez-Soler, J. M. (2012). Analysis and modelling of
youtube traffic. Transactions on Emerging Telecom-
munications Technologies, 23(4):360–377.
Behrad, A., Salehpour, M., Ghaderian, M., Saiedi, M.,
and Barati, M. N. (2012). Content-based obscene
video recognition by combining 3d spatiotemporal
and motion-based features. EURASIP Journal on Im-
age and Video Processing, 2012(1):1–17.
Briassouli, A. and Ahuja, N. (2007). Extraction and
analysis of multiple periodic motions in video se-
quences. Pattern Analysis and Machine Intelligence,
IEEE Transactions on, 29(7):1244–1261.
Carlsson, N., Dán, G., Mahanti, A., and Arlitt, M. (2012).
A longitudinal characterization of local and global bit-
torrent workload dynamics. In Passive and Active
Measurement, pages 252–262. Springer.
de Castro Polastro, M. and da Silva Eleuterio, P. M. (2012).
A statistical approach for identifying videos of child
pornography at crime scenes. In Availability, Reliabil-
ity and Security (ARES), 2012 Seventh International
Conference on, pages 604–612. IEEE.
Endeshaw, T., Garcia, J., and Jakobsson, A. (2008). Clas-
sification of indecent videos by low complexity repet-
itive motion detection. In Applied Imagery Pattern
Recognition Workshop, 2008. AIPR ’08. 37th IEEE,
pages 1–7.
Garcia, J. (2014). An evaluation of side-information as-
sisted forensic hash matching. In Computer Software
and Applications Conference Workshops (COMP-
SACW), 2014 IEEE 38th International, pages 331–
336.
Jung, S., Youn, J., and Sull, S. (2014). A real-time system
for detecting indecent videos based on spatiotemporal
patterns. Consumer Electronics, IEEE Transactions
on, 60(4):696–701.
Liu, F. and Picard, R. W. (1998). Finding periodicity in
space and time. In Computer Vision, 1998. Sixth In-
ternational Conference on, pages 376–383. IEEE.
Niyogi, S. A. and Adelson, E. H. (1994). Analyzing gait
with spatiotemporal surfaces. In Motion of Non-Rigid
and Articulated Objects, 1994., Proceedings of the
1994 IEEE Workshop on, pages 64–69. IEEE.
Schulze, C., Henter, D., Borth, D., and Dengel, A. (2014).
Automatic detection of csa media by multi-modal fea-
ture fusion for law enforcement support. In Pro-
ceedings of International Conference on Multimedia
Retrieval, ICMR ’14, pages 353:353–353:360, New
York, NY, USA. ACM.
Ulges, A., Schulze, C., Borth, D., and Stahl, A. (2012).
Pornography detection in video benefits (a lot) from
a multi-modal approach. In Proceedings of the 2012
ACM international workshop on Audio and multime-
dia methods for large-scale video analysis, pages 21–
26. ACM.
Viola, P. and Jones, M. (2001). Rapid object detection using
a boosted cascade of simple features. In Computer Vi-
sion and Pattern Recognition, 2001. CVPR 2001. Pro-
ceedings of the 2001 IEEE Computer Society Confer-
ence on, volume 1, pages I–511–I–518 vol.1.
SECRYPT2015-InternationalConferenceonSecurityandCryptography
426
