Counter-act against GAN-based attacks: A collaborative learning approach for anti-forensic detection

Kutub Uddin

Abstract:

The massive success of deep learning allows us to forge images in more perfect manners for ethical and
even unethical purposes. Several forensic methods have been proposed to expose artifacts in fake images.
However, the practice of anti-forensics (AF), particularly deep learning-based AF on digital images, has made
such forgeries difficult to detect. Therefore, a counter-AF (CAF) algorithm is necessary to reveal AF traces and
ensure the authenticity of image content. In this study, we propose a novel data-driven approach to counteract
generative adversarial network (GAN)-based AF attacks. We consider different forgery techniques, such as noise
addition, filtering, and deepfake generation to generate fake images. Subsequently, GAN-based AF attacks were
applied to conceal the forgery fingerprints such that they can deceive forensic methods. We built a new CAF
method that allows collaborative learning to detect GAN-based AF attacks. We designed a novel CAF-GAN
model by considering the commonly used GAN architectures. The proposed CAF-GAN model generates a new
image from the input image, which helps collaborative learning to detect AF images. GAN-based AF attacks can
effectively hide forgery fingerprints and significantly reduce the performance of forensic methods. However, the
proposed CAF method can effectively detect AF images in match and mismatch scenarios of AF and CAF-GAN
models.