Research: Evaluating the Robustness of Watermarking and Forensic Techniques for Detecti...

Research question: How effective and robust are current watermarking and forensic methods for distinguishing AI-generated text and images from human-created content?

Background & Motivation: The proliferation of AI-generated text and images, powered by large-scale foundation models, poses new challenges for authenticity verification in digital media. Detecting synthetic content is critical for combating misinformation, protecting intellectual property, and maintaining trust in digital communications.

Research Gap & Question: While watermarking and forensic analysis have been proposed as solutions, their effectiveness under adversarial conditions, model updates, and real-world noise is insufficiently studied across both text and image modalities. There is a need for a systematic evaluation of these detection approaches, especially as generative models continue to evolve and evade detection.

Approach & Expected Contribution: This project will conduct a comparative experimental evaluation of leading watermarking schemes (e.g., probabilistic token watermarking for text, image pixel watermarking) and forensic detection algorithms (e.g., stylometric analysis, deepfake detectors) using benchmark datasets. The study will assess detection success rates under benign and adversarial scenarios, quantify robustness to common attacks, and provide guidelines for best practice in real-world deployment.

Why it Matters: The findings will inform policymakers, AI developers, and digital platform operators about the practical reliability of current detection technologies, contributing to the responsible use and governance of generative AI systems.