FOODER: Real-time Facial Authentication and Expression Recognition

AI summary

Overview

• Research area: Computer Vision / radar-based sensing — specifically out-of-distribution (OOD) detection applied to facial authentication and facial expression recognition.
• Technical level: Advanced. The abstract assumes familiarity with FMCW radar representations (range-Doppler, micro range-Doppler), OOD detection metrics (AUROC, FPR95), encoder-decoder architectures, and vision transformers.
• Scope in one sentence: The paper introduces FOODER, a privacy-preserving radar framework that authenticates a single enrolled user by treating all other faces as out-of-distribution, then recognizes that user's facial expressions using specialized lightweight networks.

What This Paper Is About

Most facial authentication and expression recognition systems rely on cameras, which raises privacy concerns and can fail in poor lighting or when faces are occluded. This paper asks whether low-cost radar can do the job instead, using radio reflections off the face rather than images. The core challenge is that a system must not only recognize an authorized person but also reliably reject everyone else — including people it has never seen — which is framed here as an out-of-distribution detection problem.

Key Contributions

1. A radar-only, privacy-preserving facial authentication framework built on low-cost frequency-modulated continuous-wave (FMCW) radar, using both range-Doppler and micro range-Doppler representations to avoid capturing any visual data.
2. An OOD-based authentication module with a multi-encoder multi-decoder architecture incorporating Body Part (BP) and Intermediate Linear Encoder-Decoder (ILED) components, designed to classify one enrolled individual as in-distribution while flagging all other faces as out-of-distribution.
3. A hierarchical, gated expression recognition pipeline that activates only after successful authentication: a ResNet block first separates dynamic from static expressions, after which two specialized MobileViT networks classify dynamic expressions (smile, shock) and static expressions (neutral, anger) respectively.
4. A real-time evaluation on a self-collected 60 GHz short-range FMCW radar dataset, reporting 94.13% AUROC and 18.12% FPR95 for authentication and 94.70% average expression recognition accuracy, with claimed improvements over state-of-the-art OOD detection methods and several transformer-based architectures.

Main Findings

• Authentication performance: On the authors' 60 GHz short-range FMCW radar dataset, FOODER reaches an AUROC of 94.13% and an FPR95 of 18.12%. The abstract does not report additional detection metrics or per-class breakdowns.
• Expression recognition performance: The system achieves an average expression recognition accuracy of 94.70% across the expression categories it handles (dynamic: smile, shock; static: neutral, anger). The abstract does not give per-expression accuracies or confusion details.
• Competitive positioning: The authors state that FOODER outperforms state-of-the-art OOD detection methods and several transformer-based architectures. The specific baselines, margins of improvement, and comparison protocol are not specified in the abstract.
• Real-time operation: The framework is claimed to operate efficiently in real time, though the abstract provides no latency, throughput, or hardware-resource figures to quantify this.
• Hierarchical design is effective: Splitting expression recognition into a dynamic-vs-static routing stage followed by two specialized MobileViT classifiers is presented as the mechanism enabling fine-grained recognition alongside robust authentication.
• Privacy by construction: Because the system relies exclusively on radar data, no facial imagery is captured or stored, which the authors present as an inherent privacy advantage over camera-based approaches.

Methodology in Plain English

The researchers used a cheap 60 GHz radar sensor that bounces radio waves off a person's face and records how those reflections change over time and distance. Two complementary views of that signal are extracted: one that captures where reflections come from and how fast they are moving (range-Doppler), and a finer-grained version sensitive to small facial movements (micro range-Doppler).

For authentication, the system is trained to recognize one enrolled person. Rather than trying to learn what every possible other person looks like — which is impossible — it learns to treat "not the enrolled person" as out-of-distribution. The architecture uses multiple encoders and decoders, with components named Body Part (BP) and Intermediate Linear Encoder-Decoder (ILED), to make this in-distribution vs. out-of-distribution decision.

Only if authentication succeeds does the second stage run. The radar representations are concatenated and passed through a ResNet block that decides whether the expression is dynamic (in motion, like a smile or a shocked expression) or static (like neutral or angry). Based on that decision, one of two small, efficient MobileViT networks makes the final expression classification. The authors collected their own dataset with their radar setup and report authentication and expression accuracy numbers from that dataset.

Why This Matters

Impact on research: The paper pushes OOD detection into a new sensing modality — radar-based face analysis — and shows that a one-class authentication framing (treating everyone else as out-of-distribution) can be combined with downstream expression recognition in a single real-time pipeline. It also adds to the growing body of work on privacy-preserving sensing alternatives to cameras.

Real-world applications:

• Secure access control: unlocking doors, devices, or restricted areas without capturing or storing a camera image of the user.
• Driver and occupant monitoring: detecting a driver's emotional or attentional state inside a vehicle, where cameras may be unreliable at night or obstructed, and where passengers may object to being filmed.
• Human–computer interaction: expression-aware interfaces for gaming, assistive technology, or accessibility, running on low-cost radar hardware.
• Privacy-sensitive environments: healthcare rooms, bathrooms, or workplaces where continuous video surveillance is unacceptable but user verification and state monitoring are still needed.

Industry relevance: The use of commodity 60 GHz FMCW radar modules — the same class of sensor used in automotive and consumer electronics — suggests a path to deployment on existing low-cost hardware. Combining authentication with expression recognition in one radar pipeline is attractive to automotive, smart-home, and consumer-device manufacturers, though the abstract does not discuss cost, power, or integration specifics.

Future Directions

• Scaling beyond a single enrolled user: The framework currently authenticates one enrolled individual; extending it to multi-user enrollment and larger populations is an open question the abstract does not address.
• Broadening the expression set: Only four expressions are handled (smile, shock, neutral, anger), split into dynamic and static groups. Whether the hierarchical design extends to subtler or more numerous expressions is untested here.
• Generalization across sensors and settings: The evaluation uses one dataset from one 60 GHz short-range radar setup. Cross-device, cross-environment, and cross-population robustness remains unverified, and the abstract reports no such experiments.
• Quantifying the real-time claim: The paper asserts real-time operation but the abstract gives no latency, frame-rate, or compute figures, leaving the practical efficiency on embedded hardware an open question.
• Adversarial and spoofing robustness: As with any biometric system, whether radar-based authentication can be fooled by replay, mimicry, or deliberate interference is not discussed in the abstract.

Target Audience

This paper is best suited to researchers and engineers working on OOD detection, radar-based sensing, or privacy-preserving biometrics, as well as practitioners building authentication or affect-sensing systems for automotive, smart-home, or embedded platforms. Readers without a background in radar signal representations or one-class/OOD learning will find the technical framing demanding, though the high-level problem and claimed results are accessible to a broader computer vision audience.

Authors’ abstract

Out-of-distribution (OOD) detection is essential for the safe deployment of neural networks, as it enables the identification of samples outside the training domain. We present FOODER, a real-time, privacy-preserving radar-based framework that integrates OOD-based facial authentication with facial expression recognition. FOODER operates using low-cost frequency-modulated continuous-wave (FMCW) radar and exploits both range-Doppler and micro range-Doppler representations. The authentication module employs a multi-encoder multi-decoder architecture with Body Part (BP) and Intermediate Linear Encoder-Decoder (ILED) components to classify a single enrolled individual as in-distribution while detecting all other faces as OOD. Upon successful authentication, an expression recognition module is activated. Concatenated radar representations are processed by a ResNet block to distinguish between dynamic and static facial expressions. Based on this categorization, two specialized MobileViT networks are used to classify dynamic expressions (smile, shock) and static expressions (neutral, anger). This hierarchical design enables robust facial authentication and fine-grained expression recognition while preserving user privacy by relying exclusively on radar data. Experiments conducted on a dataset collected with a 60 GHz short-range FMCW radar demonstrate that FOODER achieves an AUROC of 94.13% and an FPR95 of 18.12% for authentication, along with an average expression recognition accuracy of 94.70%. FOODER outperforms state-of-the-art OOD detection methods and several transformer-based architectures while operating efficiently in real time.