Video Forgery Detection using Natural Scene Statistics

Video Forgery Detection using Natural Scene Statistics All video forgery detection techniques follow the same fundamental principle to prove the authenticity of a video and to highlight any alterations made to it. Many approaches for the detection of video forgery are evidently

Project Title

Video Forgery Detection using Natural Scene Statistics

Project Area of Specialization

Artificial Intelligence

Project Summary

Video Forgery Detection using Natural Scene Statistics
All video forgery detection techniques follow the same fundamental principle to prove the authenticity of a video and to highlight any alterations made to it. Many approaches for the detection of video forgery are evidently present in literature, but all of them have the same principle objective to detect and locate the forgery. One of the most primitive, yet critical type of detection technique is the Copy-Move forgery detection (CMFD). In this thesis, an in depth study has been conducted of the techniques that are already present in literature whilst highlighting the limitations and challenged faced. We have proposed a completely Blind (Passive) video forgery technique that uses multiple integrating natural scene Statistical (NSS) techniques to overcome those drawbacks and increase the efficiency of forgery detection. We have used a two-step approach for extracting essential features. Multiple NSS based techniques have been used for
feature extraction, then feature selection has been applied in order to eliminate redundant features. Hyper-parametric tuning of the Random Forest Classifier has been used for classification purpose. The F-Measure score (FM), recall, precision and accuracy are calculated for the proposed algorithms which are then compared with the algorithms already present in literature in order to make a
fair evaluation. The FM score achieved is 0.98 while the accuracy is 100 percent after k-fold cross validation. Dataset cross validation has also been done to prove that the model is entirely content independent. The datasets used include SULFA and LIVE. SULFA database has a total of 80 videos and share a 320x240 pixel resolution and a 30 fps frame-rate. The outputs show high efficiency for the detection and localization of forgery across different video formats and the model has been successful to overcome most of the limitations observed in the past approaches such as: it increases the efficiency of detection of forgeries, it is content independent, it is applicable for videos with static as well as moving backgrounds and it also with compressed videos. Further enhancement of the deep learning and learning transfer approach is recommended for future work.Inter Frame detection and pixel-wise detection shall be done in future.

Project Objectives

• Our aim is to devise an approach that improves the accuracy of detecting tempered videos
• The aim of this project is to search for optimal features using NSS in order to differentiate between a forged and an original video.
• Our aim is to overcome maximum limitation faced by previous methods
• Our objective is to design user-friendly graphical user interface

Project Implementation Method

Forgery detection is being done in many ways but
this paper used NSS based feature extraction techniques that is assumed to work on given video contents.Videos arrangement have done started with 10 original videos (record and compressed with either MJPEG or H264
varying on low leveled device). forged regions are generated in each original video and saved for used as a whole 80 video sequenced dataset and forged region is varying from original ones is due to their temporal pixel. We will also pursue towards using LIVE dataset cross validate with SULFA dataset to compare the results. For frame segregation we know that the picture in series combine to form a video and these pictures known as frames and they are arranged in time at particular frequency the combination of picture called group of pictures in professional terms. For our modeling we have to extract the frames fromthe video and use it for analysis. Each video made up of different number of frames and highlighting the minor detail between the frames. Frame extraction is done using MATLAB.
Next for feature extraction we are using different NSS based technique such as BRISQUE, NIQE, FRIQUEE, HIGRADE, OGIQ, SSEQ, V-BLIINDS, DIIVINE, BVQA. All these techniques used on given dataset and calculate individual feature as BRISQUE computed 36 features. These techniques are based on no reference algorithm and it do
not rely on DCT and wavelet transformation. That is highly efficient and that is useful for perceptually optimizing image processing algorithms such as denoising. These techniques depend on natural scene statistic model parameters and calculate feature using these features.

Classification is done using support vector classifier along with decision trees and random forest classifier. In SVC Regularization (C) and gamma parameter is set according to accuracy. We can have small margin place with small c and large hyperplane with large c. for gamma factor small value of gamma does not consider all supports vectors while large value of gamma considers all support vector points. The method does not produce accurate results so we move to decision tree classifier. Decision tree algorithm is used supervised learning mainly used for classification.it look
like a tree like structure the tree. Random forest classifier additional class to decision tree classifier it uses multiple decision trees at one time it is similar to ensemble learning. it will give accurate result on our datasets.

Benefits of the Project

With the modern development in modern-day era, you may nevertheless without difficulty manage a virtual image or video the use of pc software program application or a cell software program. The motive of enhancingvisible media will be as easy as to appearance appropriate in advance than
sharing to the social networking internet website’s or can be as malicious as to defame or damage one’s recognition inside the real international through such morphed visible imagery. Identity theft is one of the examples wherein
one’s identification get stolen with the aid of a few impersonators who can get right of access to the private and financial information of an innocent person. To avoid such drastic situations, law enforcement authorities have
to use a few automated device and strategies to find out whether someone is innocent or the offender. One maximum crucial question that arises here is how and what components of visible imagery can be manipulated
or edited. The answer to this question is important to differentiate the right photographs/films from the doctored multimedia.

Technical Details of Final Deliverable

The natural scene statistic model also evaluates distortion in the video like in BRISQUE Now along with GGD the paper purposes AGGD parameter for distortion specific as it is not enough to capture the whole spectrum with only GGD parameter and for AGGD we also take zero mode condition because MSCN distribution is symmetric 18 features at each scale means 36 feature is extracted as a whole to deal with distortion with GGD and AGGD fit as well as NIQE used GGD and AGGD parameter and the gaussian parameters
fit is applied to NSS based extracted features. This model also assumes image luminance with mean and deviation normalization. feature extraction method used in NIQE is similar as prior technique called BRISQUE but only difference is that it used multivariate gaussian parameter along
with GGD and AGGD parameter to compute feature of natural images which results in NSS feature and BRISQUE is distortion specific while NIQE is not. FRIQUEE based on collecting or capturing bundles of features together and map it this approach currently estimated a large set of heterogenous features which is used for perception or image quality predictions or use to model natural statistics of video frames for prediction perspective total of 561 features is in use which will be map to the model. HIGRADE will calculate total of 18 feature. F1 and f2 is computed when MSCN is fitted to GGD parameters then from f3 to f16 the features is
calculated when log derivative of luminance image is computed and then it fits to GGD parameters lastly f17 and f18 is computed using variance field. SSEQ method proposes 12 different types of features. Local entropy
(state of intensity level of frame) has been used as local pixel exhibits some statical structure and these structures are distorted in the presence of any distortion.so the paper contributed towards behavior of local pixel values
in the presence of different distortions such as sharpness, blur and noise. As state of intensity level of image for forged frames exhibits mean and skewness which is disrupted in case of distortion. DIIVINE uses mixture of gaussian parameter model is used to make a framework of neighbor pixels.Vector from feature vector is known as mixture of gaussian scale used zero mean and normalized variance vector and is used as random vector and this feature vector model used with statical characteristic of pristine
images. The paper uses the counterparts of natural images as distorted images and model them to extract the features.

Classification is done using support vector classifier along with decision trees and random forest classifier Decision tree
algorithm is used supervised learning mainly used for classification.it look like a tree like structure the tree. Random forest classifier additional class to decision tree classifier it uses multiple decision trees at one time it is similar to ensemble learning. it will give accurate result on our datasets.

Final Deliverable of the Project

Software System

Core Industry

Security

Other Industries

IT , Legal , Media

Core Technology

Artificial Intelligence(AI)

Other Technologies

NeuroTech, Big Data

Sustainable Development Goals

Peace and Justice Strong Institutions, Partnerships to achieve the Goal

Required Resources

If you need this project, please contact me on contact@adikhanofficial.com