DIF-V: A Diverse, Inclusive Synthetic Face Dataset for Fairer Verification

Face checks power everything from phone unlocks to banking logins — but training data often skews toward certain races and genders. This paper presents a method to generate high-quality, ID‑photo‑style synthetic faces that better reflect real‑world diversity, and introduces a new benchmark: Diverse and Inclusive Faces for Verification (DIF‑V).

• DIF‑V includes 27,780 images across 926 identities for fairer evaluation.
• Popular verification models still favor some genders and races.
• Applying identity style changes reduces verification accuracy.
• Generation rules enforce ID‑photo standards (lighting, pose, background).

The aim: more inclusive, transparent, and reliable face verification — with fewer privacy risks than using real faces. Learn more: https://arxiv.org/abs/2511.17393v1

Paper: https://arxiv.org/abs/2511.17393v1

Register: https://www.AiFeta.com

AI FaceVerification Biometrics FairML Bias SyntheticData ComputerVision Dataset AIethics ResponsibleAI