DFF-2025 at ACM Multimedia!

Excited to share that our paper, “What if Retrieval Could Work Before Decoding? The Case of JPEG AI Latents for Deepfake Source Attribution,” has been accepted to DFF-2025 at ACM Multimedia!

We show how latent tensors from JPEG AI can be directly used to discriminate image semantics and perform deepfake source attribution, without any training, just image-to-image retrieval. Results are promising and open up new directions for image forensics.

Thanks to my coauthors Lorenzo Catania, Francesco Guarnera, Dario Allegra, Prof Sebastiano Battiato