Comparison of Liveness Detection Methods
1. Overview of FPT AI eKYC Liveness Detection Solution
eKYC (Electronic Know Your Customer) Liveness Detection is a solution for customer identity verification using facial recognition technology and liveness detection (verifying the authenticity of a real person) to ensure that the person performing the transaction is genuine and not a spoof or using deepfake technology.
Currently, FPT AI eKYC supports liveness detection on both static images and video, with a variety of authentication methods suitable for different user experience requirements and risk appetites for each business scenario. Details of the authentication methods are presented in section 3.
2. Advantages of the FPT AI eKYC Solution
-
SDK integrates more than 30 AI models, with intuitive illustrated instructions to help users clearly understand the required steps, providing essential basic information (MRZ position, NFC chip placement, etc.), increasing the likelihood of successful authentication on the first attempt.
-
Ensures authentication steps are performed quickly (within about 4-5 seconds) so users do not have to wait long, thereby reducing the drop-off rate during the process.
-
Integrates technology for automatic detection and adjustment of lighting and camera angle to assist users when the surrounding environment is not ideal, ensuring more accurate results and optimizing request costs for banks.
-
Simplifies head movement actions in the liveness flow for money transfer transactions, making it easier for users to perform and increasing the rate of successful authentication on the first try.
-
Regularly and continuously updates AI models and new eKYC methods.
3. Detailed Comparison of Methods
To get a more intuitive understanding of how each method in the table works, please refer to the demonstration videos in Demonstration of Liveness Check Methods
| Criteria | Method 1: 03 images (left/right/front) | Method 2: Passive image | Method 3: Active image | Method 4: Flash Liveness |
|---|---|---|---|---|
| User action description | - Customer positions face within the frame - Turns to 3 positions: Front - Left - Right | - Customer positions face within the frame - Faces forward | - Customer positions face within the frame - Moves: Far - Near | - Customer positions face within the frame - Faces forward - Screen flashes red |
| Number of user actions | 03 | 01 | 02 | 01 |
| Successful recognition rate* | > 99% | > 99% | > 99% | > 99% |
| Time required** | Capture time: ~5s Processing time: ~1s | Capture time: ~3s Processing time: ~1s | Capture time: ~7s Processing time: ~1s | Capture time: ~3s Processing time: ~1s |
| Total time | ~6s | ~4s | ~8s | ~4s |
| Spoofing resistance | - Static images, printed images, screenshots: Yes - Deepfake: Yes | - Static images, printed images, screenshots: Yes - Deepfake: Yes | - Static images, printed images, screenshots: Yes - Deepfake: Yes | - Static images, printed images, screenshots: Yes - Deepfake: Better |
| FAR (False Acceptance Rate)*** | 0.133% | 0.198% | 0.175% | 0.137% |
| Image/Video size per session | 0.5 MB/image 1.5 MB/3 images | 1 MB/2 images (uploaded to server) | 1 MB/2 images (uploaded to server) | 1 MB/2 images (uploaded to server) |
| Advantages | - Highest accuracy and spoofing resistance among all methods - Fast authentication time; - High accuracy, good spoofing resistance; - Moderate data size. | - Fast authentication time; - High accuracy, good spoofing resistance; - Moderate data size; - Customer does not need to move head much. | - High accuracy, good spoofing resistance; - Moderate data size. | - Fast authentication time; - High accuracy, good spoofing resistance; - Moderate data size; - Customer does not need to move head much. - Works accurately in dark or low-light environments, etc. |
| Disadvantages | Requires users to perform more actions, and actions must be performed accurately. | - Longer execution and processing time. - Requires more user actions than method 3 but fewer than method 2. | The sudden increase in screen brightness and red flashing may be less user-friendly. | |
| Evaluation & Recommendation | - Suitable for business scenarios with high risk requirements, but low customer usage frequency. - Example: device change, onboarding, high-limit money transfer, etc. | - Suitable for the majority of banking business scenarios, and for frequent customer use to ensure processing speed and customer experience. - Example: low & medium-limit money transfer requests, etc. | - Suitable for business scenarios with high risk requirements, but low customer usage frequency. - Example: device change, onboarding, high-limit money transfer, etc. | - Suitable for most business scenarios and industries. - Ensures high security and safety requirements. - Especially optimized for end users in dark or low-light environments, etc. |
*Recognition rate tested under good conditions and successful authentication flow. As all methods share a core model, the successful recognition rate is relatively similar.
**Capture time: under ideal conditions, with users following instructions correctly.
***Results based on a dataset of 26,319 images. The number of false acceptances for each method is 35, 52, 46, and 36 respectively. False recognitions are mainly due to poor lighting, glare from environmental lights, blurry images due to movement/shaking, etc.