Image Forgery Detection Using Noise and Edge Weighted Local Texture Features

Image forgery detection is important for sensitive domains such as courts of law. The main challenge is to develop a robust model that is sensitive to tampering traces. Existing techniques perform well on a limited dataset but do not generalize well across the datasets. Moreover, these techniques cannot reliably detect tampering that distorts the texture pattern of the image. The noise patterns remain consistent throughout a digital image if its contents are not altered. Based on this hypothesis, a robust descriptor FFT-DRLBP (Fast Fourier Transformation - Discriminative Robust Local Binary Patterns) is introduced, which first estimates noise patterns using FFT and encodes the discrepancies in noise patterns using DRLBP. Features extracted are passed to Support Vector Machine (SVM) for deciding whether the image is authentic or tampered. Intensive experiments are performed on benchmark datasets to validate the performance of the method. It achieved an accuracy of 99.21% on the combination of two challenging datasets. The comparison shows that it outperforms state-of-the-art methods and is vigorous to image forgery attacks even in the presence of various post-processing operations. The performance of the method is also validated using cross-dataset experiments, which ensures its robustness and generalization.