Abstract
Biometric authentication has prospered due to its convenience and inherent security. However, early generations of biometric mechanisms often suffer from spoofing attacks. Recently, unobservable physiological signals—such as Electroencephalogram (EEG), Photoplethysmogram (PPG), and Electrocardiogram (ECG)—have emerged as potential remedies. Specifically, PPG measures blood flow changes via optical methods. While clinically used for monitoring oxygen saturation and heart rate, PPG signals contain rich individual cardiac information suitable for cybersecurity. This survey explores the unique advantages of PPG—simple acquisition, difficulty to steal, and inherent liveness detection—while systematizing the current research landscape into a unified framework and identifying future directions.
Introduction: The Vulnerability of Traditional Biometrics
Traditional biometrics like fingerprints and facial recognition face a significant "publicity" challenge: they are easily observable and replicable. Fingerprints are left on physical surfaces, and faces are captured by ubiquitous cameras, making them vulnerable to presentation attacks using 3D masks or lifted prints.
Physiological signals offer a more robust alternative. PPG, the signal used by smartwatches to track heart rate, is generated internally and responds directly to the human heartbeat. This makes it exceptionally difficult to "steal" from a distance and provides built-in liveness detection. Despite its potential, PPG research has remained fragmented. This paper proposes a standard framework to unify the field and bridge the gap toward mainstream adoption.
The Four-Layered PPG Authentication Framework
To categorize the existing body of research, we propose a four-layered architecture that tracks the signal from the sensor to the final authentication decision.
1. Signal Acquisition Layer
This layer distinguishes between contact-based methods (e.g., pulse oximeters and wearables) and remote methods (rPPG). While contact methods provide high-fidelity data, rPPG uses standard cameras to detect subtle skin color changes from a distance, offering a "frictionless" user experience.
2. Signal Conditioning Layer
The primary challenge in PPG processing is noise, particularly Motion Artifacts (MA) and baseline wander caused by respiration. Researchers employ various denoising techniques, including Kalman filters and Empirical Mode Decomposition, to isolate the clean cardiac signal.
3. PPG Representation Layer
Raw waveforms must be converted into a "digital signature." This involves extracting fiducial points (such as systolic peaks or the dicrotic notch) or applying mathematical transforms (Fourier or Wavelet) to analyze the signal's frequency components.
4. Security Authentication Layer
The final layer involves the decision-making engine. Approaches range from traditional similarity metrics, such as Euclidean distance and bit-wise comparisons, to modern deep learning architectures like CNNs and LSTMs that capture complex temporal patterns.
Dataset Analysis
A meta-analysis of over 20 public datasets reveals a diverse landscape. Key resources include PPG-DaLiA, which captures signals during daily activities, and MIMIC-III, a massive clinical dataset. However, a critical gap remains: few datasets are designed for long-term authentication, which is necessary to test how a user's signature holds up over weeks or months.
Key Insights and Performance
- The Liveness Advantage: It is nearly impossible to simulate the complex hemodynamic response of a living human using a static mask or screen, making PPG highly resistant to DeepFakes.
- The rPPG Paradox: While remote PPG is convenient, it introduces the risk of "long-distance" biometric theft via high-definition video.
- Accuracy: Many studies report Equal Error Rates (EER) below 5%, with deep learning models achieving near-perfect identification in controlled, single-session environments.
Challenges and Future Work
The most significant hurdle is the conflict between permanence and physiology. A PPG signal changes based on stress, exercise, or temperature. Future research must focus on "cross-session" authentication—identifying a user despite these physiological shifts. Additionally, as we move toward AI-driven models, we must defend against adversarial attacks and optimize algorithms for the resource-constrained hardware of wearable devices.
Summary updated