Digital Image Forensic based on Machine Learning approach for Forgery Detection and Localization

Machine learning for multimedia forensic is a new way of image forgery detection due to its amazing features of fast forgery detection. Compared with existing techniques of Deep Learning and Convolution Neural Network (“CNN”), machine learning improves security in the specific forged region under various test conditions. Some researchers use Support Vector Machine (“SVM”) and k-nearest neighbors (k-NN) algorithms to detect forgeries and another category uses unsupervised classification, including self-organization feature map (SOFM) and fuzzy c-means. But there occurs a need to address the detection speed improvement under the present scenario. The proposed algorithm has been developed using a machine learning approach to improve detection speed by pre-processing of feature extraction and feature reduction using “DWT” and “PCA” where data is trained by support vector machine (“SVM”) to provide quick results under various test conditions. This work specifies different image attacks like all types of geometric transformation, post-processing operations, etc., and presents efficiency in forgery detection and localization in case of multiple forgeries.


Introduction
With the growth of Digital transformation, multimedia security and its originality is the major concern and issue for right message communication. Multimedia forensic is a new research area for providing authentic documents with the help of powerful machine learning tools and deep learning architecture. These tools provide more security-oriented applications as they are inherent to image attacks. In machine learning, many techniques are proposed for protecting learning systems including deep learning and neural network learning by focusing on image manipulation detection. The main focus is on secure machine learning-based forensics using the training and testing process. Forensic analysis is done via Support Vector Machine and some transformation is applied to achieve robust results for feature selection to train the classifiers. The next work is carried on method for forensic analysis as any processing performed on image leaves specific traces which helps to do forensic analysis by acquisition, coding, and processing. Deep Learning suffers from many limitations that restrict its application in the security of image forensic. Training of dataset is considered to give robust results by focusing on both support vector machine and Convolution Neural Network ("CNN") classifiers. "CNN" is the programming model that helps the computer to learn from observational data. It is a class of deep neural networks that efficiently addresses various image processing tasks like pattern recognition but it is computationally a complex model as it is interconnected with large neurons. "CNN" features extraction is a data-driven process and has image classification as shape filters and a large dataset is required for the training and testing process. Figure 1 presents the simplest workflow consisting of hidden convolution parts for feature extraction and fully connected parts for classification. Machine learning gives efficient performance but has drawbacks particularly in deep learning which suffers from security in terms of large data. Another limitation occurs during the test phase where output differs and there arises a need to generate a new class of machine learning forensic.   According to Categories of Machine Learning, as described in Figure2, in classification, inputs are divided into two or more classes, and the learner must produce a model that assigns unseen inputs to one or multi-lable classification and can be handled by supervised learning. In regression, the supervised problem is generated where outputs are continuous rather than discrete. In Clustering, a set of inputs is to be divided into groups. Unlike in classification, the groups are not known beforehand, thus making this typically an unsupervised task. Density estimation finds the distribution of inputs in some space. Dimensionality reduction simplifies inputs by mapping them into a lower-dimensional space. Topic modeling is a related problem, where the program is given a list of human language documents and is tasked to find out which documents cover similar topics.
According to Classifier Evaluation Process, as described in Figure3, the complete flow is presented where the choice of the learner is the major concern point to train the dataset for feature selections and further error measurements were tested through different strategies to evaluate classification. Figure1 shows an example of image forgeries.

Basic machine Learning Concept
Machine learning is the learning distribution from input data. It is utilized to learn from experience E concerning some class of task T and performance measure P, if its performance at the task in T, as measured by P, improves with experience E. It is broadly divided into two classes that-is supervised learning and unsupervised learning. Supervised learning is the latest research area and has practical applications like classification, Pattern recognition in computer vision, etc. It builds a model based on input data for which true class is known which is sampled from input data (Labeled data/ Training data). It is also a data-driven process and the model will be only as good or as bad as the data we have. This means we cannot consider the data set of cap images and expect to use it to classify caps and rats. In this case, linear regression cannot be used to train a model on a dataset that does not have a linear correlation.
Machine learning under classification analysis task generates a statistical model with specific deficiencies like over-fitting and under-fitting. In the case of over-fitting, it occurs due to over-trained input data. This situation occurs as too many features are taken for input data and not enough data has been supplied. Another case is about under-fitting, and it occurs due to few features that are considered for training data thereby generating low and unreliable predictions. During training and testing, the learning procedure set of available data is divided into a training set for model training and a test set used for performance measuring. Under training, data validation is set to train the behavior concerning unseen data and then optimized for the choice of internal parameters of the algorithm.

"SVM"
Support Vector Machine ("SVM") is the main tool of machine learning. "SVM" is designed for binary classification and multi-class classification. The advantages of "SVM"s include high accuracy. It works efficiently in practice and has been remarkably successful in such diverse fields as natural language categorization, bioinformatics, and computer vision. It also has tunable parameters and training optimization. It gives global and unique results by avoiding the convergence to local minima exhibited by other statistical learning systems, such as neural networks. The main aim is to find a hyperplane that has the maximum margin, i.e. a maximum distance between data from different classes which predicts that future data points will be correctly classified with high confidence. Feature extraction and feature reduction are extracted as parameters in the learning of the machine. The number of extracted features was reduced.

Proposed Algorithmic Frame Work
Proposed work done in initial phase of image pre-processing and feature extraction with feature reduction process, performed efficiently to achieve the best feature to train the data set using "SVM". Analysis is done through performance matrices which defines logical and mathematical designed construct to measure how close are the actual results from what has been expected or predicted. Error measurements are the main evaluation framework in this field. Initially, threshold segmentation is used for image segmentation based on a threshold value to turn a gray-scale image into a binary image and to simplify it in a more meaningful and easier analysis. Each of the pixels in a region is similar concerning some characteristics such as color, intensity, or texture. Adjacent regions are significantly different for the same characteristic. Feature Extraction analysis is done by Fourier transform which converts a time-domain signal into constituent sinusoids of different frequencies but is limited to discarding the time information of the signal. Thus, the quality of the classification decreases as time information is lost. Feature Reduction Excessive features increase computation time and storage memory. Furthermore, they sometimes make classification more complicated, which is called the curse of dimensionality. It is required to reduce the number of features. "PCA" is an efficient tool to reduce the dimension of a data set consisting of a large number of interrelated variables while retaining most of the variations. It is achieved by transforming the data set into a new set of ordered data according to their requirements. Table2

Conclusion
In recent years, researchers have proposed a lot of approaches for forgery detection, which falls into two categories where Supervised learning provides better results than unsupervised classifiers in terms of classification accuracy. Supervised classification methods used with "SVM"s are state-of-the-art classification methods based on machine learning theory. Compared with other methods such as artificial neural network, decision tree, and Bayesian network, "SVM"s have significant advantages of high accuracy as it does not need a large number of training samples to avoid over-fitting. This work presents the image forgery detection technique using a machine learning approach to improve detection speed by pre-processing of feature extraction and feature reduction using "DWT" and "PCA" where data is trained by support vector machine ("SVM") to provide results in 25 seconds.

Acknowledgment
Thanks to the reviewers to share valuable comments to improve the quality. "Video frame copymove forgery detection based On cellular automata and local binary patterns" "Cellular automata and local binary patterns" "Copy-move region detected" "Highly efficient" 2014 16. "Speeding-up SIFT based copy-move forgery Detection using level set approach" "SIFT" "Copy move region is detected" "Less efficient" 2014 17. "Shape-based copymove forgery detection using level set approach" "Level set approach" "Copy-move region "Detection of copymove forgery using krawtchouk moment" "Krawtchouk moment" "Copy-move region detected" "Works well if the image is noisy or blurred" 2013 28. "Detection of copymove forgery using wavelet Decomposition" "wavelet Decomposition" "Copy-move region detected" "Accuracy is high" 2013 29. "Copy-move image forgery detection using local binary pattern andNeighborhoodcluste ring" "Local binary pattern and Neighborhood clustering" "Copy-move region detected" "Highly accurate" 2013 30. "Copy-move forgery detection in images via 2D-Fourier transform" "2D-Fourier transform" "Copy-move region detected accurately" "This work Detects multiple CMF and it also robust to jpeg Compression attacks even if the quality factor is lower than 50 hence highly accurate" 2013 31. "Copy move image forgery detection using mutual Information" "Mutual information" "Copy-move region detected" "Less accurate" 2013 32. "Copy move image forgery detection method using Steerable pyramid transform and texture descriptor" "Steerable pyramid transform local binary pattern (LBP)., and texture descriptor" "Copy move region is detected" "Accuracy is high" 2013 33. "Copy move forgery detection using DWT and SIFT features" "DWT" and "SIFT" "Copy move region is detected" "Defects false results also" 2013 34. "A fast "DCT" based method for copymoveforgeryDetect ion" "DCT" "Copy-move region is detected" "Will not work in the noisy image" "Polar harmonic transform(PHT)" "Scheme can detect the copy-move forgery When the copied region is rotated before being pasted." "Scheme is not efficient for scaling, local bending in images" 2012 41. "An evaluation of popular copy-move forgery detection approaches" "DCT", "DWT", K" PCA", "PCA" "Copy move region detected" "low computational load and good performance" 2012 42. "A fast image copymove forgery detection method using phase correlation" "Phase correlation" "Copy move region detected" "method is valid in detecting the image region Duplication and quite robust to additive noise and blurring" 2012 43. "Image copy-move forgery detection based on crossing shadow division" "DWT" and crossing shadow "Copy-move region detected" "Algorithm has advantages of low Computational complexity" 2011 44. "Detection of copycreate image forgery using Luminance level techniques" "Luminance level techniques" "Copy-create image forgery" "Time-consuming and less accurate" 2011 45. "Detecting copy-paste forgeries using transform-invariant features" "Transform-invariant features" "Copy-paste forgery detection" "Difficult detection in case of a blurred image" 2011 46. "Detecting copy-move forgery using nonnegative matrix factorization" "Non-negative matrix factorization" "Copy-move region is detected" "Some geometric distortions (e.g. rotation, Reflection, etc.) Can render the method invalid"