4 ms·

#2 is called "watermarking". If the watermarking is done for each recipient differently, then it is called "traitor tracing". #1 and #2 will eventually be defe

#2 is called "watermarking". If the watermarking is done for each recipient differently, then it is called "traitor tracing". #1 and #2 will eventually be defeated by newer models as they get trained on the protected images or need to pass some output condition. You should know that these so-called protective image distortions are the kind of thing that the generative adversary in GAN training learns to do and classifier learns to ignore. Only #2 and #4, watermarking and signing are likely to be continue to be effective as models improve. For content signing there is the C2PA specification https://c2pa.org/specifications/specifications/2.2/index.html https://c2pa.org/specifications/specifications/2.2/index.htm... that describes how to include signatures in most media types.

rijal028 1y ago
Thank you for this incredibly valuable and insightful feedback. You've perfectly articulated the core challenges and long-term viability of the different defensive layers. I completely agree that the "AI Poison" layer is likely a temporary defense. Your validation of watermarking and especially cryptographic signing as the most robust long-term solutions is a crucial takeaway for me. And thank you especially for pointing me to the C2PA specification. I was not aware of this standard, and aligning my concept with it is a critical next step. I will be studying this document carefully. This is exactly the kind of expert feedback I was hoping to get by sharing my work here. I truly appreciate you taking the time.