Table of Contents
Recently, there has been a notable trend in the financial and security fields towards establishing secure authentication systems using face recognition for self-verification. Consequently, the significance of systems that prevent fraudulent authentication through the use of others' faces or manipulation is highly emphasized.
iBeta's Presentation Attack Detection (PAD) is a comprehensive test that evaluates the effectiveness of face spoofing detection. It meticulously assesses the system's performance, ensuring precise and consistent operations in real-life scenarios. The test duration is limited to 8 hours and employs easily accessible materials like paper and electronic devices found in typical households and office environments as spoofing samples. The test alternates between 150 spoofing attacks and 50 real faces for thorough comparison. To successfully pass the test, achieving a 0% success rate in spoofing attacks and a matching rate is required. The evaluation method is divided into three stages.
This test evaluates the performance of detecting various forms of face spoofing. It utilizes two essential metrics, APCER (Attack Presentation Classification Error Rate), and BPCER (Bona Fide Presentation Classification Error Rate), to measure the spoofing detection system's effectiveness. APCER represents the error rate of classifying spoofed faces as genuine faces, while BPCER indicates the error rate of classifying genuine faces as spoofs. Lower error rates for both metrics signify more accurate face spoofing detection performance.
In this test, the spoofing detection system captures and processes facial data while conducting quality checks. The performance of the system's data collection is verified through the evaluation of algorithms that recognize faces and extract facial features from the data.
This test evaluates the overall performance and biometric comparison capability of the spoofing detection system. It involves comparing real individuals with spoofed data and verifying performance using three metrics: Imposter Attack Presentation Match Rate (IAPMR), False Non-Match Rate (FNMR), and False Match Rate (FMR).
IAPMR represents the match rate between spoofed and genuine face data, measuring the degree of similarity between fake and real face data.
FNMR is the error rate of incorrectly identifying the same person as a different person, while FMR is the error rate of incorrectly recognizing a different person as the same person. Lower values for these three metrics indicate higher face spoofing detection performance and greater identity authentication accuracy.
ALCHERA’s ongoing efforts aim to improve forgery detection performance using the RGB camera. Compared to other types, RGB cameras present greater difficulty in identifying forgeries. However, their low hardware constraints open up possibilities for widespread technological applications.
RGB cameras capture and analyze facial images using color information, but they have limited access to thermal data, detailed facial features, and depth information, which are essential for effective forgery detection. Furthermore, they are significantly affected by lighting conditions, making data augmentation and algorithm development crucial for enhancement.
To ensure consistent recognition and detection performance in real-life situations, the training incorporates data reflecting various lighting environments. Among the brightness adjustment data augmentation techniques, three common ones are Channel Shuffle, Random Brightness, and Hue-Saturation adjustments. Among these techniques, 'Random Brightness' adjustment was selected and applied during the training process, especially for the environment where the RGB camera is primarily utilized.
In real-life scenarios, capturing videos with devices equipped with RGB cameras makes it challenging to eliminate focus and shaking issues. To overcome this, artificial blur is introduced and used during the training process. Various types of blurs, such as Gaussian Blur, Median Blur, and Motion Blur, are employed. Among these options, ALCHERA specifically utilizes Motion Blur. This technique effectively compensates for the phenomenon caused by object movement during shooting, making it highly practical for real-world applications.
In a self-conducted test in January of this year, ALCHERA demonstrated exceptional performance by achieving 100% forgery detection and boasting over 99% accuracy in real-face usability tests. The test utilized various samples, including regular print outputs, high-quality outputs, 2D facial masks, and 3D facial models, and was conducted passively without requiring user participation. ALCHERA showcased approximately 35 to 39 times higher performance compared to two domestic face recognition AI companies.
Leveraging its high usability and versatile technology, ALCHERA provides non-face-to-face ID verification AI solutions across diverse fields. A notable application is in the financial sector, where high-performance verification technology is crucial in preventing identity theft and fraudulent transactions. ALCHERA's solution is integrated into Toss Bank, one of the leading internet-specialized banks in South Korea, enabling facial authentication for services such as self-camera login, identity verification, and account opening, with built-in facial forgery detection technology.
Through ALCHERA's Face Recognition AI, Toss Bank streamlined authentication procedures that previously required ID submissions or video calls to a simple facial verification process, allowing accurate identity authentication with just a selfie.
From the iBeta PAD testing stage to the present, we have closely examined facial forgery detection performance. ALCHERA continues to surpass the limitations of hardware and software through continuous performance improvement, further expanding the application of face recognition AI across various domains.
[Article] The Importance of iBeta’s Passive Liveness Detection Testing - Part 1