Active Liveness Detection: Defending Against Deepfakes in Video Verification
Modern generative models produce highly realistic fake selfies and documents that easily bypass standard face matching. High-quality synthetic images can fool basic liveness checks, leading to successful attacks on verification systems. To counter this, active liveness detection is being deployed, using video-based dynamic prompts.
The method requires users to follow a randomized sequence of head movements: up, down, left, and right. A blink prompt is added when the head returns to the center. Each action is captured within a strict timeframe, confirming a live person is present in front of the camera.
During the process, the system captures a clear frame with the user’s eyes open and head properly aligned. This single frame is then used for face matching against the ID document, ensuring high-accuracy comparison without requiring multiple retakes.
Smartphone Implementation
The workflow is seamlessly integrated into mobile apps: users simply scan a QR code to activate the camera. The system captures the ID document and runs video verification in a single, continuous session.
The randomized prompt sequence effectively blocks pre-recorded video attacks. The system analyzes movement patterns, facial depth, and skin texture to distinguish live captures from deepfakes.
To meet stringent face matching requirements, a custom-trained model is deployed. It achieves baseline metrics of AUC ROC ≥ 0.9966 and Accuracy ≥ 0.9808. The model is trained on up-to-date datasets with rigorous quality control.
Additional Verification Options
- Cross-referencing with government databases and official registries.
- Adjustable verification strictness levels.
- Seamless integration with existing KYC/AML platforms.
These features allow the verification process to be tailored to specific use cases, significantly reducing false positives.
The approach drastically cuts verification time: optimal head positioning and lighting ensure first-attempt success. It effectively eliminates static spoofing attempts, including AI-generated faces and forged documents.
Performance Metrics & Efficiency
| Metric | Value |
|------------|------------|
| AUC ROC | 0.9966 |
| Accuracy | 0.9808 |
These results were validated across a wide range of devices and lighting conditions. The system remains highly resilient against attacks leveraging GANs and video editing software.
Key Takeaways
- Active liveness detection relies on randomized movement prompts to prevent pre-recorded video spoofing.
- Capturing a frame at the optimal angle simplifies face matching and boosts accuracy.
- The custom model delivers an AUC ROC of 0.9966 and an accuracy rate of 0.9808.
- Smartphone integration via QR code requires no additional software.
- Database cross-referencing options provide an extra layer of security.
The solution has been approved by clients with strict anti-fraud requirements. It integrates smoothly into production environments with minimal API adjustments.
— Editorial Team
No comments yet.