On September 29, WuHan Homsh Technology Co., Ltd. and Wuhan University of Science and Technology officially signed the technical development contract Development and Edge-Side Adaptation of Iris Presentation Attack Detection (PAD) Algorithm. Over an 8-month period, the two parties will jointly develop an iris anti-spoofing algorithm that requires only a single frame of near-infrared eye image and no additional hardware, and deploy it directly on Homsh’s C20 iris terminals and MD31 iris recognition modules.
Accurate recognition is only half of iris recognition; the other half is first confirming that what is in front of the camera is a real human eye.
Iris texture is stable and hard to replicate, making it one of the most secure biometric traits recognized worldwide. But “hard to replicate” does not mean “no one tries”: printed iris photos, textured colored contact lenses, 3D prosthetic eyes, and eye videos replayed on screens are all counterfeits that attackers may hold up to the camera. In the industry, such attacks are called “presentation attacks”, and the technology to detect them is known as PAD.
The difficulty lies in the constant evolution of attack methods. Many existing solutions perform well against attack types seen in training, but their detection rates drop significantly when encountering new contact lens materials, new types of prosthetic eyes, or higher-definition screens. Other solutions rely on multispectral illumination, structured light, or multi-frame video to determine liveness — while effective, they also drive up device costs and recognition time. On top of this, stricter anti-spoofing measures lead to higher false rejection rates for real users, and high-precision models are difficult to fit into embedded chips. These contradictions are exactly the hard nuts this cooperation aims to crack.
The division of labor for this cooperation is clear. The research team of Wuhan University of Science and Technology is responsible for the independent research and implementation of the algorithm, including the construction of attack sample datasets, design of discrimination models, as well as model pruning, quantization and edge-side adaptation. Homsh provides C20 iris terminals and MD31 iris recognition modules, takes charge of device integration and debugging, and organizes two levels of acceptance testing.
Universities excel at algorithm research, while enterprises provide real devices and application scenarios, with two-level acceptance checks to ensure the quality of deliverables. 4 key attack types, 2 deployment platforms, 8-month R&D cycle.
The project sets itself a strict prerequisite: using only a single 640×480 monocular near-infrared grayscale eye image, with no additional hardware, to identify multiple types of attacks including printed irises, textured contact lenses, 3D prosthetic eyes and screen replay. Meanwhile, it aims to maintain a high attack detection rate while strictly controlling the false rejection rate for real users.
Around this goal, the two parties have defined four R&D directions: characterizing iris texture, edge structure and specular reflection simultaneously with multi-scale features to capture subtle differences in material and imaging between real eyes and counterfeits; adopting leave-one-attack-type-out training and open-set discrimination to improve adaptability to “unseen attacks”; integrating quantization-aware training and structured pruning into the training process to ensure the model remains accurate when compressed onto embedded chips; and finally, ensuring credible evaluation from the data source.
Four R&D directions covering the full chain of features, generalization, deployment and data.
Iris information is sensitive personal information. Every real human subject in the project will be individually informed of the purpose of the data, storage period and their rights, and will sign a written informed consent form on an individual basis, with the right to withdraw at any time. All data is stored encrypted throughout the process, with access logs kept for all operations.
Evaluation also follows strict rules: samples are grouped by subject and attack source, and samples from the same source will not appear in both the training set and test set at the same time. The independent test set is sealed and kept by Homsh, and self-test results from the R&D party cannot substitute for acceptance testing. Performance is judged solely on unseen data.
The algorithm will ultimately run on two platforms: the PC side for use with Homsh C20 iris terminals; and the RK3588 platform as the main controller for MD31 iris recognition modules, with the algorithm running directly inside the module. The project progresses according to four milestones: the mid-term acceptance assesses algorithm accuracy, while the final acceptance assesses pruning and quantization, dual-platform deployment and real device integration.
Eight months, four milestones, each with clear deliverables and confirmation methods.
From self-developed iris recognition chips to embedding the full recognition pipeline into edge devices, Homsh has always been committed to one mission: making identity authentication more trustworthy. Anti-spoofing capability is the most critical part of “trustworthiness”.
This collaboration with Wuhan University of Science and Technology brings together the research strength of universities and the product scenarios of enterprises. We look forward to delivering a self-controllable iris anti-spoofing algorithm that can be directly integrated into terminals in eight months’ time. We also welcome more universities and research teams to communicate and cooperate with Homsh.
