A low-light robustness framework for face verification with enhancement-based recovery

Facial verification is one of the cornerstones of modern biometric security systems. However, its robustness in realistic low-light conditions is severely under-investigated. Here, we propose a systematic framework for low-light robustness, which explicitly models the photometric degradation and integrates the enhancement-based preprocessing before the embedding extraction. We evaluate three state-of-the-art face verification backbones, AdaFace, MagFace and Dblib under five levels of illumination degradation, including deterministic attenuation (Level 1-3), a stochastic random variant and a novel Random Extreme setup that simultaneously applies nonlinear gamma compression, spatially varying luminance suppression and multiplicative sensor noise to simulate realistic nighttime surveillance conditions. To alleviate the latter performance degradation, we propose a hybrid enhancement strategy, which sequentially integrates Gamma correction, Linear intensity scaling and Contrast-Limited Adaptive Histogram Equalisation (CLAHE) to recover discriminative facial features before the embedding extraction. Evaluated across four public datasets LFW, AgeDB-30, CALFW and CPLFW, and show that severe illumination degradation can lead to sensitive models losing accuracy down to 85.75%, while the proposed hybrid strategy consistently brings performance back to over 98%. The hybrid approach also outperforms state-of-the-art flow-based enhancement methods (LLFlow, UPT-Flow and JTE-CFlow) in a comparative evaluation, showing its superior stability, especially when paired with classical metric-learning backbones. Our results confirm the inherent photometric robustness of margin-based embedding representations and the necessity to systematically incorporate illumination-aware preprocessing in face verification pipelines for reliable deployment under real-world low-light conditions.