DiffusionPrint: Learning Generative Fingerprints for Diffusion-Based Inpainting Localization

DiffusionPrint addresses a critical gap in digital forensics: standard Image Forgery Localization (IFL) relies on PRNU (Photo-Response Non-Uniformity) and camera sensor noise, which are destroyed when a latent diffusion model (LDM) re-decodes an image through a VAE. The project's defensibility is low (3) because it is currently a paper-driven reference implementation with zero stars and no community traction; the 'moat' is purely academic and easily replicated once the weights or methodology are public. Frontier risk is medium because while labs like OpenAI/Google focus on proactive watermarking (e.g., SynthID), they are less likely to build adversarial forensic tools for third-party models. However, platform domination risk is high from companies like Adobe or specialized forensic firms (e.g., Reality Defender, Sentinel) who could integrate these specific LDM-aware spectral analysis techniques into their existing suites. The displacement horizon is short (1-2 years) because as diffusion architectures evolve (e.g., moving away from standard VAEs to DiTs or different latent spaces), the specific 'fingerprints' identified here will require constant retraining and adaptation, a classic cat-and-mouse game in forensics.