Lightweight Test-Time Adaptation for Robust Out-of-Distribution Face Recognition in Web Services

Abstract

Face recognition systems have the potential to support diverse services in Web 3.0 applications, yet two critical challenges remain underexplored. First, existing benchmark datasets are demographically biased and underrepresent elderly East Asian users, limiting fair and inclusive deployment. Second, sensor noise, lighting shifts, and motion blur introduce out-of-distribution (OOD) corruptions that cause severe accuracy degradation and undermine reliability in decentralized environments. To address these issues, we introduce the Korean Senior Face Benchmark, consisting of 700 images of 70 Korean senior celebrities, enabling realistic assessment for an underrepresented demographic. We quantitatively demonstrate that recent state-of-the-art models suffer significant performance drops under realistic corruption conditions, highlighting the need for enhanced robustness. Finally, we show that a lightweight test-time adaptation (TTA) strategy can recover OOD performance without retraining, making it well-suited for edge devices and distributed infrastructures while preserving user privacy. Experiments show accuracy gains of up to 41.5% under the most severe corruptions, along with improvements in intra-class compactness and inter-class separability in the embedding space. The proposed benchmark and adaptation pipeline lay a practical foundation for building distributed, fair, and privacy-aware face-recognition services in Web 3.0 applications.