✳flâneur — a map of the web's best reading

Proteus

github.com · 547 words · saved by 1 readers

Cryptographic image provenance research based on perceptual hashing as a more robust replacement to watermarks. More info at https://proteus.photos! Or check out our Twitter threads. Advancements in AI (like @OpenAI's Sora!) are making it increasingly difficult to tell if content on the internet is real or AI-generated. Social media cracked mass content distribution, and AI is cracking content creation at scale. So how do you recognize AI images/videos vs. camera content vs. edited media? Previous efforts have focused on embedding watermarks—the white house recently announced related plans. However, existing methods to embed watermarks are not resilient to how content changes when transmitted (e.g., compression, crops, screenshots). If you embed a watermark into an image, multiple transformations to that image will fundamentally remove the watermark. But interestingly, an image's perceptual representation varies much less amidst these same transformations. For example, while the pixels

Proteus Photos: Robust Cryptographic Image Provenance Cryptographic image provenance research based on perceptual hashing as a more robust replacement to watermarks. More info at https://proteus.photos ! Or check out our Twitter threads . Introduction Advancements in AI (like @OpenAI's Sora!) are making it increasingly difficult to tell if content on the internet is real or AI-generated. Social media cracked mass content distribution, and AI is cracking content creation at scale. So how do you recognize AI images/videos vs. camera content vs. edited media? Difficulties with Watermarking Previo

Explore this link on the map →

related reading