Retrosynthesis of Synthetic Media for Explainable AI Provenance Forensics

Tracing AI-Generated Content Without Modifying Generators: The paper addresses a practical challenge in responsible AI by proposing a method to identify whether images came from specific generative models without requiring changes to the generators themselves—important for real-world deployment where modifying existing systems isn't feasible.

Round-Trip Verification as a Forensic Tool: The framework uses a "round-trip consistency" approach where images are encoded, processed by a generator, and then decoded back to verify their origin. If the decoded image matches the original input, it confirms the image came from that specific generator, providing a practical forensic method for detecting synthetic media.

Explainability Over Black-Box Detection: Unlike many AI detection systems that simply output "real or fake," this approach provides interpretable evidence showing why an image is identified as synthetic—offering transparency that's crucial for responsible AI deployment and public trust in content verification systems.

Maintaining Visual Quality While Enabling Traceability: The method preserves the visual quality of generated images while embedding traceability information, addressing a key tension in responsible AI: enabling beneficial generative AI applications while maintaining accountability for synthetic content creation.

Practical Solution for MLaaS Platforms: By working within "fixed-generator settings" without requiring watermarks or architectural modifications, this approach is designed to work with existing commercial generative AI services, making it immediately applicable to real-world content verification challenges.