Uncertainty-Aware Face Anti-Spoofing with Epistemic and Aleatoric Referral Under Unseen Attacks

Face anti-spoofing, or face presentation attack detection (PAD), must often decide on attack types absent from training. Existing uncertainty-aware PAD methods use a single confidence or uncertainty score, which cannot show whether an unreliable decision comes from an unfamiliar attack or a poor-quality capture. This study proposes a lightweight Bayesian framework on MobileNetV3-Large that combines Monte Carlo dropout with a heteroscedastic logit head to estimate epistemic and aleatoric uncertainty separately, and uses both in an ACCEPT, REJECT or REFER rule with thresholds fixed on validation data containing only known attacks. Unseen attacks, defined as attack types excluded from training and validation, were tested on SiW Protocol III in both directions and on SiW-Mv2 Protocol II, which holds out each spoof type in turn. Over five seeds, the complete configuration reduced the average classification error rate (ACER) from 4.01% for the deterministic model to 2.61% at 97.09% coverage, with ACER of 2.91% and 3.46% on the two Protocol III directions and 2.68% on SiW-Mv2. Two re-implemented baselines reached lower ACER (1.31% and 1.94%) but do not separate uncertainty sources. Expected calibration error remained high (0.2983). The contribution is reason-coded referral rather than lower error.