Resource-aware test-time adaptation for mobile face anti-spoofing under SWAP constraints

Face anti-spoofing (FAS) models for mobile authentication face a difficult deployment gap: they must remain robust under domain shift (camera, illumination, attack medium) while balancing biometric accuracy against strict speed, model-weight, and power-consumption limits (SWAP). Existing test-time methods improve cross-domain accuracy but ignore on-device resource limits and, more critically, can catastrophically degrade when source and target domains differ substantially. We propose a resource-aware test-time adaptation (RA-TTA) framework that updates only normalization affine parameters, a compact classifier head, and class prototypes, and only when (a) drift is detected against calibrated source statistics, (b) pseudo-labels pass a confidence-and-augmentation reliability gate, and (c) a token-bucket budget controller permits adaptation under hard speed and power caps. Evaluated on OULU-NPU and Replay-Attack datasets with a deliberately challenging single-source 1 M-parameter MobileNetV3-Small backbone, RA-TTA delivers a 25% relative ACER reduction (4.83% vs 6.47%) on intra-domain data where adaptation safely fires, while preserving source-model behaviour under extreme cross-domain shift (43.70% on OULU to Replay-Attack, identical to No-TTA) where Tent collapses by 5.7% ACER. A drift-threshold sensitivity sweep validates the calibration heuristic. The runtime profile is deployment-class – 1.6 ms/3.04 mJ per frame on Samsung Galaxy S25, 3.83 ms/3.95 mJ on Galaxy A56, and 5.94 ms/6.27 mJ on entry-level Galaxy A17, with a 3.84 MB ONNX footprint – but cross-domain ACER at this scale is below production thresholds for security-critical face authentication. The contribution is the safety mechanism, not state of the art accuracy. Reaching production accuracy on the same backbone requires multi-source training, which is orthogonal to the adaptation-safety question this paper addresses.