Enhancing Image Attribution via Unforgeable Fingerprints With Sublinear Communication

Summary

The paper proposes a proactive scheme for tracing deepfake images back to their generative source models, using fingerprints embedded in the model's training images. Cryptographic commitments during training and fingerprint embedding make the original fingerprint undeniable, and polynomial identity testing verifies it lightly. Verifying a 100-bit fingerprint needs only 30 to 40 bits of communication, a 60 to 70 percent overhead reduction, with communication complexity reduced to O(log L), where L is the fingerprint length.