Image Forensic Using Machine Learning

Abstract

Nowadays, it is challenging to trust any digital image due to the convenient availability of manipulation software like Photoshop, GIMP, and Coral Draw etc. Therefore, it becomes tough to differentiate between an authentic image and tampered image. Traditional methods for image forgery detection generally use handcrafted features. The challenge with the traditional image tampering detection approaches is that most of the methods need improvement as only certain features are identified. These days, Machine learning (ML) and deep learning (DL) are widely used in image forgery. These techniques prove their efficacy with better accuracy and other performance parameters than traditional methods. There are many types of image forgery, like copy-move, splicing, and retouching. In this thesis, copy-move and splicing forgery are detected using ML and DL techniques.The first algorithm provides a copy-move image forgery detection using machine learning and deep learning. In this work, machine and deep learning algorithms are proposed to find out different image forgeries. First, the proposed algorithm applies color illumination in preprocessing, then Scale Invarient Feature Transform (SIFT) is used to extract features, and Support Vector Machine (SVM) classifies correct forged pixels. The proposed methodology gives better results for CMF detection as Precision=97.25%, Recall=100%, and F1=98.53%.The second algorithm provides a deep convolution neural network (DCNN) that uses automatic feature extraction and localizes copy-move forgery and splicing forgery. In the feature extraction and localize forgery, the performance can be enhanced using the ML and DL. Finally, the applications of proposed color illumination, convolution neural network, and semantic segmentation are demonstrated for forgery detection. The proposed algorithm performance accuracy is calculated on the CASIA v1.0 validation set, and the test set is 98% and 99%, respectively. The performance accuracy is calculated on the CASIA v2.0 validation set, and the test set is 98% and 98%, respectively. The DVMM dataset forgery detection accuracy is 97%. The BSDS300 dataset forgery detection accuracy is 98%. The proposed algorithm is tested on imagelevel on CMFD dataset and achieved performance accuracy, i.e. Precision (P) = 98%, Recall (R) = 100% and F1 = 99%.The third algorithm presented robustness of algorithms against geometrical attacks using color illumination, a deep convolution neural network, SIFT, and SVM. Geometrical attacks, such as scaling, rotation, and JPEG, were identified. The plain CMF attack detection results are: P=97.25%; R=100% and F1=98.53%. The JPEG CMF attack detection results are: P=71.44%; R=58.44% and F1=63.77%. The scale CMF attack detection results are: P=85.2%; R=74.8% and F1=79.1%. The rotation CMF attack results are: P=87.83%; R=76.33% and F1=86.16%. Comparison with state-of-the-art techniques proves the efficacy of the presented algorithms. In the future, suggested algorithms can be implemented on real-time applications with some improvements.