Image Forgery Detection using Deep Learning

Image Forgery Detection using Deep Learning

Aishwarya Sedamkar1, Sachin Deshpande2

1 PG Student, Department of Computer Engineering, Vidyalankar Institute of Technology

2Professor, Department of Computer Engineering, Vidyalankar Institute of Technology

---------------------------------------------------------------------***---------------------------------------------------------------------

Abstract - Undoubtedly one of the most active study areas in the field of blind image forensics is copy-move forgery detection (CMFD). The majority of known algorithms rely on block and key-point approaches, alone or in combination. Deep convolutional neural network techniques have recently been used in picture classification, image forensics, image hashing retrieval, and other areas. These techniques have outperformed more conventional techniques in these areas. The work makes a novel convolutional neural network-based copy-move forgery detection algorithm suggestion. The suggested method makes minor adjustments to the net structure using small training samples after using an existing trained model from a sizable database like ImageNet. Results from the experiments demonstrate that the method we suggested produces a forgery image created automatically by computer with a simple copy-move operation in a satisfactory manner.

In order to automatically build hierarchical representations from the input RGB color photographs, this system proposes a new image fraud detection strategy based on deep learning that makes use of a convolutional neural network (ELA CNN) and transfer learning model. Image splicing and copy-move detection applications are the focus of the proposed ELA CNN and transfer learning methodology. The basic high-pass filter set used in the calculation of residual maps in the spatial rich model (SRM), which serves as a regularizer to effectively suppress the effect of image contents and capture the subtle artifacts introduced by the tampering operations, is initialized with the weights at the first layer of our network rather than using a random strategy. In order to extract dense features from the test images, a pre-trained model like Vgg, Densenet, or ELA CNN is used as a patch descriptor. A feature fusion strategy is then investigated in order to produce the final discriminative features for SVM classification.

Digital forensics research on forgery detection and localisation is important and has recently received more attention. Traditional approaches typically rely on manually created or shallowly learned features, but these have poor description capabilities and high computational costs. Deep neural networks have recently demonstrated their ability to efficiently learn the hierarchical representations of complicated statistical data from high-dimensional inputs. In this paper, we propose an improved mask regional convolutional neural network (Mask R-ELA CNN) that adds a Sobel filter to the mask branch in order to capture more distinguishing features between tempered and non-tempered areas. As a support job, the Sobel filter encourages predicted masks to have picture gradients that are close to those of the ground truth mask.The whole network is capable of spotting copy-move and splicing, two different types of image modification. The suggested model outperforms the current conventional methods for forgery detection, according to experimental results utilising two standard datasets (Casia) and ELA CNN.

Key Words:  forgery detection, forensic, image processing, ELA CNN, VGG, DenseNet