Q&A between researchers who are improving deepfake detection

Persona

Oct 21, 2026

In the real world, attackers don't just generate a selfie, submit it, and move on. If one attempt fails, they’ll refine the image and try again. Existing datasets often fail to account for these variations, leading the researchers’ detection models to fall short in real-world scenarios.

Zhaofeng Si, a PhD student at University at Buffalo who interned at Persona as a research scientist, and Saatvik Billa, an ML engineer and researcher at Persona, worked together this summer on a fraud taxonomy and a real-world selfie fraud benchmark.

Here they interview each other about what they built, what academic datasets miss, and whether publishing it helps attackers.

━━━━━━━━━━━━━━

EXPLORE THE DATA

Dataset (Hugging Face): https://huggingface.co/datasets/withpersona/persona-fas-bench-v1
Interactive threat taxonomy: https://threats.withpersona.com/

━━━━━━━━━━━━━━
CHAPTERS

00:00 The gap between academic benchmarks and real-world fraud

00:30 Introductions

01:03 The main takeaway for researchers

01:41 Why the taxonomy is a path, not a checklist

02:40 The most important stage: asset refinement

03:32 What academic datasets miss

04:04 How the taxonomy connects to the dataset