Published Paper Details:

Due to a multi-fold growth in the diffusion of multimedia data through the open and unprotected Internet, content authentication of digital photographs has caught the attention of forensic professionals and security researchers. Attackers who are astute come up with new approaches to test state-of-the-art forensic tools for detecting forgeries in digital photographs. On benchmarked datasets, feature engineering techniques have produced accuracy of up to 97 percent. Deep learning algorithms have showed promise in a variety of picture classification tasks, but they are unable to discover hidden patterns in digital images that can be used to consistently detect image forgeries. Deep learning techniques for forgery detection have a state-of-the-art accuracy of up to 98 percent on benchmarked datasets. The proposed approach aims to improve detection accuracy even more, bringing it close to 100%. To mine patterns responsible for accurate forgery detection, this work uses a synergy of created features based on colour attributes and deep features using the image's luminance channel. The first Stream computes 648-D Markov-based features from the image's quaternion discrete cosine transform. In the second Stream, the image's Local Binary Pattern is extracted using the YCbCr colorspace's luminance channel. Local binary feature maps are also input into the pre-trained ResNet-18 model to get a 512-D feature vector named 'ResFeats' from the model's convolutional base portion's last layer. An 1160-D feature vector is created by combining handcrafted features from Stream I and ResFeats from Stream II. The method is also tested on the CASIA v1 and CASIA v2 datasets, and classification is done with a shallow neural network. On benchmark datasets, the suggested fusion-based technique has a 99.3 percent accuracy.