FARE: A Deep Learning-Based Framework for Radar-based Face Recognition and Out-of-distribution Detection

TL;DR

FARE tackles radar-based face recognition with essential OOD detection in privacy-sensitive environments. It combines Range-Doppler Images and micro-RDIs in a dual-path network: a Primary Path (PP) optimized for ID using triplet loss $L_{PP}$ (margin $m=2$) and Intermediate Paths (IPs) as linear autoencoders trained with MAE losses $L_{IP}$ to detect OOD via reconstruction errors, with the PP frozen during IP training. Evaluation on a 60 GHz radar dataset yields an ID accuracy of 99.30% and an OOD AUROC of 96.91%, surpassing a ResNet34 baseline and several state-of-the-art OOD detectors. The approach demonstrates a practical, privacy-preserving radar-based biometric solution for smart-home deployments, leveraging intermediate feature representations to improve OOD safety. Together, these results indicate a viable path toward robust, privacy-aware biometric systems in indoor environments.

Abstract

In this work, we propose a novel pipeline for face recognition and out-of-distribution (OOD) detection using short-range FMCW radar. The proposed system utilizes Range-Doppler and micro Range-Doppler Images. The architecture features a primary path (PP) responsible for the classification of in-distribution (ID) faces, complemented by intermediate paths (IPs) dedicated to OOD detection. The network is trained in two stages: first, the PP is trained using triplet loss to optimize ID face classification. In the second stage, the PP is frozen, and the IPs-comprising simple linear autoencoder networks-are trained specifically for OOD detection. Using our dataset generated with a 60 GHz FMCW radar, our method achieves an ID classification accuracy of 99.30% and an OOD detection AUROC of 96.91%.

Figures (2)

• Figure 1: This figure summarizes our architecture. The upper yellow section (1) represents the PP with initial modality-specific feature extractor blocks (1.1) and a combined feature extractor block (1.2). Minimized versions of the PP, also with a yellow background, represent the positive and negative samples, alongside the anchor for triplet training. A KNN classifier is used after extracting embeddings from the PP. The lower pink section (2) shows the intermediate paths (IPs). The PP remains frozen, as indicated in the figure, with training focused solely on the IPs. The IPs, which are linear encoder-decoder architectures, are located at the end of each layer of the PP and are responsible for OOD detection. The section shown in white and labeled 3 represents the navigator.
• Figure 2: Confusion matrix to show the classification performance of FARE.