Improving Detection of Unprecedented Anti-forensics Attacks on Sensor Pattern Noises Through Generative Adversarial Networks

Yijun Quan, Chang-Tsun Li

Abstract: Sensor Pattern Noise (SPN) based methods have been widely applied in multimedia forensics. With SPN being a noise-like signal in an image, anti-forensic attacks can be performed by applying different manipulations to attenuate or remove the SPN and mislead the investigators. Thus, forensic investigators should identify the images subject to such manipulations before applying SPN-based methods. Despite neural network-based classifiers having shown their strength in detecting specific anti-forensics attacks, such a classifier may not generalise well for other manipulations. Given that various manipulations can attenuate or remove SPNs and a classifier’s training set may only include certain types of manipulations, a classifier may encounter images subject to attacks unprecedented to it and unable to distinguish them from the pristine ones. To address this problem, we propose a training strategy using Generative Adversarial Networks (GAN). This strategy shifts the classifier’s excessive emphasis on the manipulation-specific features with the resultant classifier generalising better for unprecedented anti-forensics attacks.