Greater Bay Area university team solves face recognition fraud problem, equipping financial anti-fraud with "Hawkeye" vision

From large-scale smart city construction to mobile application login, face recognition, as a flagship technology of artificial intelligence, has quickly permeated various fields such as financial payments, public security, government services, and commercial marketing due to its convenience and efficiency, once becoming a “standard configuration” for digital transformation.

However, as “face-based withdrawals” and “face payments” rapidly become popular, financial terminal identity verification faces unprecedented security challenges. Facial images can be high-fidelity restored from paper photos, electronic screens, 3D head models, and even DeepFake videos. Traditional recognition systems are easily penetrated, leading to a high incidence of financial fraud cases caused by “face attacks,” severely undermining the trust foundation of intelligent financial systems.

In response to the current mainstream artificial intelligence forgery attacks in the market, Professor Yu Zitong’s team from Greater Bay Area University utilizes key large model technologies, intelligent image acquisition and imaging techniques, multi-modal heterogeneous biometric anti-spoofing and anti-fraud technologies to develop more intelligent, universal, and highly secure face liveness detection technology. They have further developed modules equipped with algorithm models, forming electronic and intelligent hardware and software systems capable of recognizing live faces within seconds and providing corresponding analysis reports with an accuracy rate of over 99.9%.

Yu Zitong is a tenured associate professor at Greater Bay Area University, engaged in research on micro-visual computing and multi-modal foundational models. He told Southern Finance reporters that initially, face recognition technology was widely used in mobile phone unlocking, attendance, and security, with relatively limited application scenarios. After earning his master’s degree, Yu Zitong worked at a security company for a year, gaining a deep understanding of face recognition technology in frontline security applications, and identifying shortcomings and pain points in algorithms and model architectures.

During his studies in Finland and Singapore, Yu Zitong’s team proposed an spatial domain weak signal enhancement algorithm based on central difference convolution operators, providing a new approach to improve model robustness against high-fidelity face attacks and environmental variations, successfully applying it to face anti-fraud fields.

In Yu Zitong’s view, current face recognition technology is becoming widespread, with frequent security issues such as personal privacy leaks and AI forgery attacks. Face liveness detection has become the most critical and challenging part of intelligent financial terminals, serving as the technological foundation for safeguarding national financial security.

With China’s comprehensive acceleration of smart finance construction, the banking industry is rapidly shifting from traditional outlets to intelligent terminals such as ATMs, remote video tellers (VTM), and smart tellers (STM). Intelligent financial terminals have become an essential part of national financial infrastructure, directly affecting the inclusiveness and convenience of financial services, and are crucial for national financial and information security stability, serving as the first line of defense against systemic risks and ensuring transaction trustworthiness.

“Currently, liveness detection relies on rapid development of deep learning, but key aspects such as high-security algorithms, cross-ethnicity robust models, and physiological signal-level detection are still dominated by Europe and the United States. There are obvious gaps domestically, including difficulties in recognizing high-fidelity fake samples, poor generalization across ethnicities, limited performance in cross-modal fusion, and challenges in algorithm engineering deployment,” said Yu Zitong.

He further explained that currently, AI face swapping and 3D printing technologies make forged faces nearly indistinguishable from real ones. Traditional texture and optical flow features have high false detection rates, making it difficult to meet financial-grade security thresholds. Domestic algorithms are mostly trained on single skin tones and scenarios, with significantly reduced recognition accuracy for dark skin and overseas users, limiting international deployment. Existing solutions mainly rely on appearance cues from RGB, infrared, and depth images, with low utilization of non-contact physiological cues, and lack model interpretability and robustness in complex environments. Anti-fraud algorithms are mostly confined to academic or software levels, without forming independent intellectual property or mass-producible “financial-grade integrated devices.”

Facing the trend of internationalization and intelligence in smart finance, there is an urgent need to break through these bottlenecks and build an independent, controllable, secure, and trustworthy face anti-fraud technology system that can be used across regions.

As a national high ground for financial equipment manufacturing and AI innovation, Guangdong has, over the past five years, taken face recognition and anti-fraud technology as engines, ushering in a new wave of fintech development opportunities.

Focusing on “key technologies for face anti-fraud and the development of high-security intelligent financial terminals,” Yu Zitong’s team from Greater Bay Area University is conducting systematic research in areas such as spatiotemporal weak signal enhancement, multi-modal cross-domain adaptation, and physiological signal-level liveness detection. They aim to achieve high-precision liveness detection and identity verification in complex environments and across different populations, creating a new generation of internationally leading, exportable intelligent anti-fraud terminals.

In the field of liveness detection, Yu Zitong’s team has proposed a multi-view collaborative perception technology based on physiological signals, integrating contactless physiological signals with appearance features. This method overcomes the limitations of traditional liveness detection relying on single RGB modality and being easily affected by environmental factors, significantly improving detection capabilities and interpretability against complex scenarios and high-fidelity attacks. By synchronously capturing facial dynamic information with multiple cameras, extracting physiological signals such as heart rate, blood flow rhythm, and blood oxygen changes (rPPG), and combining dynamic behavioral features like micro-expressions and blinking frequency, they achieve multi-layered liveness judgment from appearance to physiological levels.

“We developed a technology similar to a 3D cloud platform in 2024, which enhances face capture and interaction convenience. Currently, our face anti-fraud technology is at the forefront domestically, demonstrating excellent reliability and real-time performance, with good user interaction and international adaptability, capable of solving cross-ethnicity and cross-race challenges,” Yu Zitong said.

In fact, the intelligent financial anti-fraud terminal system developed by Yu Zitong’s team integrates and industrializes algorithm research, building a fully controllable “algorithm-hardware-system” face anti-fraud smart financial terminal. Through soft and hardware co-optimization, they have developed embedded AI acceleration modules to enable real-time detection and edge inference of anti-fraud algorithms. The system integrates modules such as face recognition, anti-fraud detection, identity verification, and encrypted communication, forming a mass-producible financial-grade anti-fraud intelligent terminal. Currently, this terminal has been deployed in core financial institutions like Industrial and Commercial Bank of China and has been successfully commercialized in Southeast Asia, with broad industry applicability and international prospects.

Yu Zitong explained that through multimodal fusion and cross-domain adaptive technological innovation, the system addresses the trustworthiness of face liveness detection in complex environments within the smart finance sector, achieving a seamless integration of algorithm innovation, engineering implementation, and industrialization. The project has been recognized by the Guangdong Society of Image and Graphics through scientific appraisal, achieving pioneering results in key multimodal face anti-fraud technology and industrialization, with independent intellectual property rights and reaching international advanced levels; notably, the multi-visual cue-driven central difference convolution model and multimodal cross-domain adaptation method for face anti-fraud are at the international forefront.

It is understood that related equipment from this project has been promoted through various financial devices of China UnionPay, with a total of 1.814 billion yuan in sales over three years and a net profit of 72.64 million yuan, achieving the top domestic market share and ranking among the top three globally.