Circumventing shortcuts in audio-visual deepfake detection datasets

Bucharest Deep Learning is back for its second session!

After our successful relaunch, we continue our series of research talks with a presentation focused on Deepfake Detection. Join us for a deep dive in the topic alongside Stefan Smeu, presenting his recent CVPR 2025 Highlight paper.

The Talk: Circumventing Shortcuts in audio-visual deepfake detection datasets with unsupervised learning.

A simple spurious cue (leading silence in fake videos) enables near-perfect separation, meaning many prior models learned artifacts instead of true manipulation signals. Leveraging audio-visual synchronization learned exclusively from real data not only mitigates dataset-specific biases, but also provides a more future-proof foundation for deepfake detection that is robust to unseen generators and evolving manipulation techniques.

Logistics:

• Date & Time: Thursday, March 26 | 18:30 - 19:30
• Location: FMI New Building (Politehnica Business Tower)
• Address: Bulevardul Iuliu Maniu, nr. 15G, Etaj 5, Room 503