MVGSig: Multi-View Graph Fusion for Out-of-Distribution Detection in FMCW Radar-Based Person Identification

Abstract

Radar-based person identification using FMCW sensors enables privacy-preserving authentication but risks overconfident predictions for out-of-distribution (OOD) individuals. We propose MVGSig, a multi-view graph fusion framework for open-set identification. Six complementary signal representations (spectral, temporal, time-frequency, wavelet, cepstral, and higher-order statistical) are extracted from a single FMCW radar sensor and fused through graph attention networks with learnable edge biases and quality-weighted pooling. View-Aware Collaborative Calibration (VACC) transfers knowledge from high-quality to low-quality views during training. A learned meta-detector combines node, edge, graph, and reconstruction anomaly scores. Under leave-one-subject-out cross-validation on a public ten-participant dataset, MVGSig reduces the false accept rate at 95% true positive rate from 0.697 to 0.375 ± 0.103, a 46.2% reduction relative to the raw signal baseline, and increases mean AUROC from 0.846 to 0.896 ± 0.031. FPR@95 is the primary metric because admitting unknown individuals is the security-critical failure. Closed-set accuracy reaches 66.5 \pm 10.1% on data collected for vital-sign estimation, a deliberately low-discriminability regime where overconfident predictions are particularly concerning. Ablation under matched scoring conditions attributes + 0.025 AUROC to multi-view fusion and another + 0.025 to collaborative view learning, which eliminates more remaining false accepts (32.1% versus 20.8%).