ACM Multimedia 2025
We are excited to share that our paper
“Towards Reliable Audio Deepfake Attribution and Model Recognition: A Multi-Level Autoencoder-Based Framework”
has been accepted for presentation at the DFF Workshop @ ACM Multimedia 2025, to be held in Dublin 
!
!In this work, we present LAVA, a multi-level architecture for audio deepfake attribution that combines a convolutional autoencoder, attention mechanisms, and open-set rejection to recognize both the technology and the specific model used to synthesize fake audio.
The framework is trained and evaluated across multiple datasets (CodecFake, ASVspoof2021, FakeOrReal) and demonstrates strong performance even under open-set conditions and distribution shifts (ASVspoof2019 LA).
Check out the architecture below!