Finding Ai Generated Deepfake Faces In The Wild

LinkedIn's March 2024 research account shows why a synthetic-face detector needs testing against generators it has never encountered. Its model learned from real profile photos and generated faces, then scored separate evaluation images for signs of an artificial face.

At a threshold that incorrectly flagged 0.5% of real faces, it detected 98% of synthetic faces from generator families represented in training. Detection fell to 84.5% for generators excluded from training. The difference matters for account-abuse detection: strong results on familiar image generators leave a larger blind spot when the source of a fake face changes.