Gaussian filtering detection based on features of residuals in image forensics
Jae Jeong Hwang, Kang Hyeon Rhee · 2016
For a design of the Gaussian filtering (GF) detection (GFD) in the altered digital images, this paper presents a new feature vector that is obtained from two residual feature kinds, which are extracted from the Gaussian filter residual (GFR) in the space domain and the frequency transform residual (FTR) in the frequency domain in an image. The feature vector of the space domain is obtained with the most ten features among the gradients of the neighbor vertical and horizontal lines of the Gaussian filtering residual. The feature vector of the frequency domain is obtained with the most ten features among the frequency coefficients of the frequency transform residual (FTR). In the proposed method, the defined 20-dim. feature vector is trained in a SVM (Support Vector Machine) classifier for the GFD of the forged images. In the experiment, the test item uses three kinds: the area under curve (AUC), the minimal average decision error, and the classification ratio. The performance is excellent both at GF (3×3) and (5×5) vs. JPEG (Quality Factor=90), Downscaling (90%) and Upscaling (110%) respectively on the proposed GFD scheme. However, the measured AUC are above 0.9 thus the grade evaluation of the proposed GFD method is rated as “Excellent (A)”.