The survey on watermarking methods for proactive defense against deepfake

Abstract

As generative models advance, deepfake content is becoming indistinguishable from reality and passive forensic detection methods are becoming increasingly ineffective. The misuse of generative tools provide for adversaries opportunities for social engineering, disinformation campaigns and fraud. This requires a new class of forensics tools based on the preemptive marking of authentic content in order to defend it from being used for deepfake media generation or disinformation campaigns. In this survey we provide a comprehensive analysis of watermarking solutions for the purpose of proactive defense from deepfake. We identified most of the existing deepfake watermarking solutions in literature and provided taxonomy for them. Also we identified core metrics and datasets for training deep learning models for proactive defense watermarking. We make quantitative and qualitative comparisons of existing solutions, their methods, metrics and purposes. In the end we provide a summary of open problems and challenges in the field. This survey lays a foundation for future development of proactive deepfake defense methods and policies for generative AI compliance.