Smooth Filtering Forensics Based On Frequency-Domain Features

Relying on the easy-to-use image editing tools,some malicious manipulations have already threatened the authenticity and integrity of images,especially the electronic evidence in the crimes.Since smooth filters are often applied to conceal the traces left by various forgeries,it is significant to achieve the blind detection of them.Detection of smooth filtering and further identifying the types and parameters of filters could recover the tampering history of an image and help determine whether an image has been forged.In this paper,we focus on three widely used spatial smooth filters,such as average,Gaussian smoothing and median filters,and proposed two methods for identification of smooth filtering types and template parameters,respectively.For the types detection,we constructed a novel low-dimensional feature vector called the annular accumulated points(AAP)based on different distribution patterns in the frequency domain of images,where a common non-linear support vector machine(SVM)is introduced to achieve the classification.For more challenging parameters identification,we proposed an approach based on the convolutional neural network(CNN).In detail,through firstly adding a transform layer,we get distinguishable frequency-domain patterns and then put them into a CNN model,to identify the template parameters of spatial smooth filters.Comparison experiments on a large composite database show that the proposed method performs outstanding among all existing methods concentrated on median filtering detection,especially in the practical scenarios of JPEG compression and low resolution.And for the types detection,our method obtained excellent accuracy about99%.Our CNN-based method could also achieve good performance than some other applicable manually extracted features,and remain effective in the scenarios of low resolution and JPEG compression.