Why academic selfie fraud benchmarks fail in the real world
Academic deepfake datasets rarely static: generate an image, label it real or fake, done. Real-world fraud doesn't hold still. Zhaofeng Si, a PhD student at University at Buffalo and research scientist intern at Persona, breaks down the gap. Public academic datasets are mostly typical face swaps and diffusion-model-generated images with a fixed label. In the real world, there's an attack-and-defense loop. Fraudsters actively probe for ways to bypass detection, so defenses have to keep adapting instead of training against a static benchmark.
