A colored contact lens printed with another person’s iris texture, a high-definition printed iris photo, a realistic prosthetic eye, or an AI-synthesized iris image — all of these share one thing in common: they can fool an iris recognition system that only focuses on pattern matching.
In the industry, such methods have a standard term: Presentation Attack. No matter how high the recognition accuracy of a system is, if it cannot distinguish between a real human eye and a counterfeit in front of the camera, its security is fundamentally compromised. The international standard ISO/IEC 30107-3 has listed Presentation Attack Detection (PAD), commonly known as liveness detection, as an independent evaluation item for biometric systems.
Homsh has recently completed the R&D and practical verification of its self-developed iris presentation attack detection system, HOMSH-PAD. Tested on an independent dataset of 18,516 images, the system achieves a Detection Equal Error Rate (D-EER) of 0.50%.
I. Verify Authenticity First, Then Confirm Identity
As an independent module placed before the recognition workflow, HOMSH-PAD operates as follows: after the device captures an eye image, the system first locates the iris and outputs an authenticity score via a deep learning network. Only images confirmed as real human eyes proceed to the recognition step; those identified as attacks are rejected immediately.
The current version covers four major types of attack methods: textured colored contact lenses, printed iris images, prosthetic eyes, and AI-synthesized irises. Among them, prosthetic eyes and printed iris images have a zero miss rate in testing; the miss rate for AI-synthesized irises stands at 0.08%. The system maintains effective detection even for new contact lens patterns never seen during training — this means the model has learned the common features of attacks, rather than simply memorizing specific patterns.
Plain-language interpretation of the test data
Out of 1,000 various attack attempts, approximately 993 are blocked;
Out of 1,000 legitimate users, approximately 3 are incorrectly rejected.
II. Behind the Numbers: A Contest of Data Discipline
The iris liveness detection field has long been plagued by a persistent issue: models deliver impressive metrics in the lab, but fail frequently when deployed in real-world scenarios. Homsh’s R&D team believes the problem usually lies not in the model itself, but in the training data — hidden statistical shortcuts accidentally correlated with true/false labels allow the model to take “shortcuts” instead of learning the actual characteristics of attacks.
To address this, the team has developed a training data domain governance methodology to systematically detect and eliminate such shortcuts before data is stored and training begins. A total of 75,919 iris images from 10,750 subjects were used for training, calibration and testing, with completely non-overlapping subject groups across the three datasets — every eye in the test set was never seen by the model during training. This methodology is the core independent innovation of the project and has been protected by patent filings.
All metrics are calculated in accordance with the ISO/IEC 30107-3 international standard. The decision threshold is determined on an independent calibration set, with no exposure to the test set throughout the process.
III. Engineering Design Built for Real-World Deployment
HOMSH-PAD is fully engineered for practical deployment:
● Lightweight: The model is approximately 30 MB in size, capable of real-time inference on a CPU, and compatible with all form factors from embedded access control terminals to servers.
● Easy to integrate: Integrators only need to input raw captured images; all preprocessing is handled internally by the SDK, enabling integration with just a few lines of code.
● Calibratable: Supports one-click threshold recalibration using real user samples from the deployment site, allowing the false rejection rate to be tailored to the actual user group.
● Fail-safe security: Abnormal inputs are classified as attacks by default, leaving no “gray area” for attackers to exploit.
IV. Completing the Final Piece of the Puzzle
While the iris recognition engine answers “who this is”, HOMSH-PAD answers “whether this is a real person”. Combined, they enable Homsh’s iris recognition solutions to fully meet the increasingly explicit liveness detection requirements of high-security scenarios such as government affairs, public security and finance.
Going forward, the team will continue to expand the attack sample library, complete multiple rounds of training stability verification, and advance full-device integration with all-in-one iris recognition products as well as third-party authoritative evaluations.
