💻 computer science

Event-based Liveness Detection using Temporal Ocular Dynamics: An Exploratory Approach

This paper explores event-based liveness detection by leveraging the microsecond-resolution temporal ocular dynamics captured by event cameras to effectively distinguish genuine eye movements from replay attacks, achieving up to 95.37% accuracy with a spiking convolutional neural network.

Original authors: Nicolas Mastropasqua, Ignacio Bugueno-Cordova, Rodrigo Verschae, Daniel Acevedo, Pablo Negri

Published 2026-04-30

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Nicolas Mastropasqua, Ignacio Bugueno-Cordova, Rodrigo Verschae, Daniel Acevedo, Pablo Negri

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to unlock your phone with your face. Usually, the camera takes a picture of you. But a clever thief could just hold up a video of your face on a tablet, fooling the camera into thinking you are there. This is called a "replay attack."

This paper proposes a new way to stop these thieves by changing the camera itself. Instead of a standard camera that takes pictures (frames) like a movie, the authors used an Event Camera.

The Camera That Sees "Changes" Instead of "Pictures"

Think of a standard camera like a flipbook. It takes a full picture, then another full picture, 30 or 60 times a second. Even if you move your eye super fast, the camera only sees you in those frozen snapshots.

An Event Camera is more like a room full of tiny, hyper-fast motion sensors. It doesn't take pictures. Instead, it only "speaks up" when something changes in a specific spot.

If a pixel stays the same color, the sensor stays silent.
If a pixel gets brighter or darker (like when your eye moves), the sensor fires a tiny "event" with a timestamp accurate to the microsecond.

It's like the difference between watching a slow-motion video of a hummingbird (standard camera) and having a microphone that only clicks every time the wings beat (event camera). The event camera captures the motion itself, not just the still images.

The "Eye Dance" Test

The researchers realized that human eyes move in very specific, fast ways called saccades (quick jumps from one point to another) and blinks. These movements happen so fast that a standard video camera often blurs them or misses the tiny details.

Because event cameras are so fast, they can see the exact "dance" of your eye muscles.

Real Humans: When you look at a target, your eye jumps instantly. The event camera sees a sharp, precise burst of activity.
Fake Videos (Replay Attacks): If a thief plays a video of your eye on a screen, the screen has to refresh its image (usually 60 times a second). This creates a "stutter" or a "jitter" in the movement. The event camera sees this as a messy, unnatural pattern because the screen can't reproduce the micro-second speed of a real eye jump.

How They Tested It

The team built a system that acts like a bouncer at a club:

The Challenge: The system asks you to do something specific, like "Look left quickly" or "Blink."
The Response: It watches your eyes with the event camera.
The Decision: It analyzes the "rhythm" of your eye movement.
- If the rhythm is smooth and micro-second precise, it's You (Genuine).
- If the rhythm has the "digital stutter" of a screen, it's a Thief (Replay Attack).

The Results

They created a new dataset by recording real people and then recording videos of those people being played back on a screen to simulate an attack. They tested their system against these fakes.

The Winner: They used a special type of AI (called a Spiking Convolutional Neural Network) that is designed to understand these "event" signals.
The Score: The system correctly identified real people vs. fake videos 95.37% of the time.
The Comparison: They also tried using standard video frames (like a normal phone camera) with a different AI, but the event camera method was more accurate.

Why This Matters (According to the Paper)

The paper argues that because event cameras capture the speed and timing of movement rather than just the look of a face, it is much harder for a thief to fake a successful attack. The "digital artifacts" (the tiny glitches) left by a screen are impossible to hide from this type of sensor.

In short: They built a security system that doesn't just check what your face looks like, but how fast and smoothly your eyes move, using a super-fast sensor that sees the world in "moments of change" rather than "pictures." This makes it very hard for a video recording to trick the system.

1. Problem Statement

Face liveness detection (Presentation Attack Detection) is critical for preventing spoofing attacks (e.g., replaying a video) in biometric systems.

Limitations of Current Methods: Traditional RGB-based methods often achieve high accuracy in controlled, intra-dataset settings but fail to generalize across different sensors, lighting conditions, and attack scenarios. They frequently rely on dataset-specific artifacts rather than intrinsic liveness cues.
The Specific Challenge: Replay attacks are particularly difficult to detect because they mimic the visual appearance of a real face. However, they struggle to reproduce the temporal dynamics of genuine human motion, specifically fast eye movements (saccades and blinks), due to temporal resampling (frame rates) and display artifacts.
The Gap: While event cameras (neuromorphic sensors) offer microsecond resolution and high dynamic range, their application to remote-eye liveness detection (where the camera is >50cm away) remains largely unexplored.

2. Methodology

The authors propose an active, event-based anti-spoofing system that leverages temporal ocular dynamics.

A. Data Collection & Dataset

Base Dataset: Extended the existing RGBE-Gaze dataset (synchronized RGB and Event camera data).
New Replay Attack Dataset: Created a custom replay attack dataset by playing genuine RGB clips on a 60Hz IPS screen and re-recording them using a multi-sensor rig (RealSense D435i, DAVIS346, and Prophesee EVK4).
Protocol: An active Challenge-Response mechanism was implemented. Users are prompted to perform rapid eye movements (saccades or blinks). The system analyzes the event stream to verify if the motion is genuine.
Scale: 17 subjects (4 female, 13 male) with synchronized RGB and Event data for both genuine and replay scenarios.

B. Event Data Processing

The system processes asynchronous event streams $(x, y, t, p)$ where $p$ is polarity.

Feature Extraction:
- Activity Profile ( $A_p$ ): A continuous estimate of event activity using exponential decay kernels to model temporal dynamics.
- Surface of Active Events (SAE): A frame-like representation encoding the recent temporal history of events in a spatial neighborhood, enabling the use of standard CNNs.
- Statistical Features: Mean event rate, polarity balance, and inter-event intervals (IEI) to detect screen refresh artifacts.

C. Detection Pipeline

The pipeline performs two concurrent tasks:

Ocular Movement Segmentation: Detecting the precise timing of blinks and saccades to validate the challenge response.
- Blink Detection: Uses signal processing on the activity profile (searching for zero-crossings in smoothed $A_{on} - A_{off}$ signals).
- Saccade Detection: Uses a Temporal Convolutional Network (TCN) with dilated convolutions to model temporal dependencies, outperforming simple peak detection.
Face Liveness Classification: Determining if the sequence is genuine or a replay.
- Model: A Spiking Convolutional Neural Network (SCNN) processes event data directly (voxel grid representation).
- Comparison: Also evaluated Vision Transformers (ViT) on SAE frames and standard CNNs.

3. Key Contributions

Novel Modality Application: First study to investigate event-based liveness detection leveraging temporal ocular dynamics in remote-eye scenarios.
New Dataset: Introduction of a replay attack dataset derived from RGBE-Gaze, specifically designed for event-based liveness research.
Temporal Segmentation Method: An event-driven method for the precise temporal segmentation of saccades and blinks in remote settings.
Robustness Analysis: Demonstration that event cameras capture intrinsic motion cues (microsecond resolution) that are inherently difficult for replay attacks to replicate due to display limitations.

4. Results

The system was evaluated on the custom dataset using a subject-independent split (80/20).

Liveness Detection:
- The Spiking Convolutional Neural Network (SCNN) achieved the best performance with 95.37% Top-1 Accuracy.
- ACER (Average Classification Error Rate): 4.65% (APCER: 4.20%, BPCER: 5.10%).
- The SCNN outperformed the Vision Transformer (ViT) trained on RGB frames (92.15% accuracy), proving the superiority of event-based temporal features for this task.
Ocular Movement Detection:
- Blink Detection: Simple signal processing achieved an F1-score of 95.35%.
- Saccade Detection: The TCN model achieved an F1-score of 89.65%, significantly outperforming the baseline peak detection method (62.72%).
Feature Analysis: Statistical analysis confirmed that replay attacks exhibit distinct artifacts, such as lower mean event rates (due to screen brightness limits) and inter-event intervals clustered around the screen's frame rate (e.g., 16.6ms for 60Hz), which are absent in genuine event streams.

5. Significance and Future Work

Robustness: The approach demonstrates that event cameras provide a robust solution against replay attacks by analyzing the physics of motion (temporal dynamics) rather than just static appearance.
Privacy & Latency: Event cameras offer low latency and privacy-preserving capabilities (no full-frame images, only sparse changes), making them ideal for real-time biometric systems.
Future Directions:
- Expanding the dataset to include diverse attack instruments (different screens, print attacks, masks, AI-generated forgeries).
- Increasing dataset scale and environmental diversity to improve generalization.
- Developing fully event-based pipelines that remove dependency on RGB frames for face registration.

In conclusion, this work establishes that temporal ocular dynamics captured by event cameras are a highly effective and robust cue for distinguishing genuine users from replay attacks, achieving state-of-the-art accuracy in a challenging remote-eye setting.