Content area
Technological advancements, combined with the widespread availability of digital images, have resulted in ubiquitous image forgery. The instances of forged images used for political propaganda, legal evidence manipulation, defamation of individuals, pornographic image creation of celebrities, and other actions that can cause social turmoil are proliferating. Verifying the integrity and authenticity of digital images is impractical for humans, necessitating the development of a robust image forgery detection model.
This thesis demonstrates methods for copy-move, splicing, and deepfake image forgery detections by combining Error Level Analysis (ELA), Discrete Cosine Transformation (DCT), or Discrete Fourier Transformation (DFT) with a Deep Learning Convolutional Neural Network. In ELA, the test image is saved at a definite compression level. Then, it is compared to the original image because the differently compressed tampered regions exhibit inconsistent error levels with the unaltered image regions. DCT is a block-based analysis that extracts DCT coefficients from spatial images, which can be used to perform image forgery detection. DFT transforms spatial images into a frequency domain, which can help to detect high-frequency-based anomalies that differentiate authentic and forged images. The experiments used the CASIA ITDE V2.0 dataset for copy-move and splicing and the 140k Real and Fake Faces dataset for deepfake image forgeries.
For the CASIA ITDE V2.0 dataset, the Deep Learning model trained on ELA images yielded a training accuracy of 97.20% with a loss of 0.07. When trained on a DCT-based model, a comparable training accuracy of 95.15% and a loss of 0.11 were achieved. Likewise, when a DFT-based model was used, the training accuracy was 95.76%, and the loss was 0.11. For the deepfake dataset, the ELA-based model attained a training accuracy of 93.02% with a loss of 0.17. However, when a DCT-based model was used for training, training accuracy plummeted to 88.14%, and loss rose to 0.28. The DFT-based model showed the least training accuracy of 81.43% and the highest loss of 0.40. These results demonstrated that ELA-based analysis performed effectively for various methods of image forgery, while the DCT-based model performed modestly.
