Linear and Non-Linear Filter-based Counter-Forensics Against Image Splicing Detection

Digital images are widely used as primary sources of evidence in today’s world, spanning security, forensics, and legal domains. However, image tampering poses a shallow technical skill barrier with the wide availability of sophisticated, easy-to-use image manipulation software. Tampered images are often used intentionally for unlawful and malicious purposes. One of the most common forms of image manipulation attack is image splicing, which is performed by combining regions from multiple source images to synthesize an artificial image that looks natural. Digital forensic measures have been widely explored in the literature to detect such type of image forgery. However, the recent growth of counter-forensics poses a threat to such forensic/security measures. Forensic techniques can be easily deceived by adopting counter-forensic manipulation of forged images. In this work, we explore different linear and non-linear filtering-based counter-forensic modifications to digital images and hence investigate the after-effects of those, in terms of severity of such manipulations in rendering state-of-the-art forensic splicing detection methods useless. In this paper, we implement two forensic image splicing detection techniques based on feature extraction from image along with machine learning and deep CNN with transfer learning. Then, different filtering techniques have been applied to the image dataset, investigating their effectiveness as a counter-forensic attack against image splicing detection. Experimental results show that the Gaussian filter and Average filter are the two most effective counter-forensic filtering methods against image splicing detection, suggesting the need for further strengthening the existing family of forensic techniques.