Maestro (Machine-learning Age Estimator: Smart, Trustworthy, Responsive, On-device): An On-Device Age Verification Pipeline with Zero Data Retention
Francesco Celino, Andrea Bricola
Age-gated applications need accurate, privacy-preserving age checks that run on-device, yet open-source age estimators remain too coarse for the adolescent band where false accepts matter most. We present Maestro (Machine-learning Age Estimator: Smart, Trustworthy, Responsive, On-device), a full verification cascade that detects faces, selects high-quality frames, rejects presentation attacks, and estimates age, returning only a binary pass/fail with zero biometric retention. The age core combines a MobileNetV4-Conv-Large backbone with a CORAL ordinal head, enabling threshold-agnostic probability queries at any legal age cutoff without retraining. On a curated adolescent benchmark (n=576), Maestro achieves a mean absolute error of 4.30 years, a 57% reduction versus MiVOLO (10.24years, p<0.001) and lower error than DeepFace and InsightFace, aiming to be the state of the art for adolescent age estimation. On a challenging held-out evaluation set, the same model reaches 2.86 years MAE in the security-critical 10–25 range. The pipeline is trained in PyTorch and targeted for on-device deployment via PyTorch Mobile. Keywords: age estimation; age verification; computer vision; on-device AI; ordinal regression; face anti-spoofing; biometric privacy; edge computing