Report

A Comprehensive Study on Third Order Statistical Features for Image Splicing Detection Xudong Zhao, Shilin Wang, Shenghong Li and Jianhua Li Shanghai Jiao Tong University, Shanghai P. R. China 1. Introduction • Digital Image Forensics: Active detection methods Watermarking, fingerprint, signature, etc. Passive detection methods Pixel based, camera based, physics based, statistical features based, etc. Latest image forgeries http://www.fourandsix.com/photo-tampering-history 2. Proposed method 2.1 Preprocessing 8*8 block DCT domain X 11 X 21 X m1 X 12 X 22 X m2 X ij U X ijU , T s X 1m X 2m X mm 1 i, j m 1 U ( n , k ) , k 0, 0 n 7 2 2 U ( n , k ) 1 cos( (2 n 1) k ), 1 k 7, 0 n 7 2 16 E h ( i , j ) X ( i , j ) X ( i 1, j ) E v ( i , j ) X ( i , j ) X ( i , j 1) 2.2 Third order statistical features States : , ,..., Conditional Co-occurrence Probability Matrix (CCPM) 1 2 N P ( 1 , 1 1 ) P ( 1 , 1 2 ) CCPM P ( 1 , 1 N ) P ( 2 , 1 1 ) P ( 2 , 1 2 ) P ( 2 , 1 N ) P ( N , N 1 ) P ( N , N 2 ) P ( N , N N ) 2nd Markov 2 nd P ( 1 1 , 1 ) P ( 2 1 , 1 ) Markov P ( N 1 , 1 ) P ( 1 1 , 2 ) P ( 2 1 , 2 ) P ( N 1 , 2 ) P ( 1 N , N ) P ( 2 N , N ) P ( N N , N ) 2nd CPM P ( 1 , 1 , 1 ) P ( 2 , 1 , 1 ) nd 2 CPM P ( N , 1 , 1 ) P ( 1 , 1 , 2 ) P ( 2 , 1 , 2 ) P ( N , 1 , 2 ) P ( 1 , N , N ) P ( 2 , N , N ) P ( N , N , N ) • Class separability, an overview • (a) (b) (c) Lda projections of (a) CCPM, (b) 2nd Markov and (c) 2nd CPM. All the samples are extracted from Columbia Image Splicing Detection Evaluation Dataset. 2.3 Feature Dimensionality Reduction Dimensionality of Proposed features N3 dimensional feature for each direction. (e.g. 7 states CCPM, there are totally 2*73 dimensional features.) PCA for Dimensionality Reduction PCA is a linear transform that maps the original features onto an orthogonal vectors spanned subspace. Coefficients and variances distributions of third order statistical features 3. Experimental Results and Performance Analysis 3.1 Image Dataset Columbia Image Splicing Detection Evaluation Dataset 3.2 Classifier Support vector machine (SVM) Radial basis function (RBF) ½ for training and the left ½ for testing Detecting accuracy is the average of 30 runs. 3.3 Detection Results Comparisons (a) (b) (a) (b) • 3.4 Robustness Test Jpeg compression Gaussian low pass filtering Image scaling Detecting results over Jpeg compressed image dataset Detecting results over Gaussian low pass filtered image dataset Detecting results over scaled image dataset 4. Conclusions Third order statistical features, more discriminative information compared with lower order features Detection performance of CCPM in Block DCT domain outperforms that of 2nd Markov and 2nd CPM PCA maps the most discriminative features onto the first several principal components, which reduce the dimensionality greatly. Robustness of both third order features and second order features will be further improved.