Iris Style Transfer: Enhancing Iris Recognition with Style Features and Privacy Preservation through Neural Style Transfer
Mengdi Wang, Efe Bozkir, Enkelejda Kasneci · 2025
Read Iris Style Transfer: Enhancing Iris Recognition with Style Features and Privacy Preservation through Neural Style Transfer by Mengdi Wang, Efe Bozkir, Enkelejda Kasneci (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-04707.txt.
Introduction
Wang, Bozkir, and Kasneci used style transfer in twenty twenty-five to lift iris ID and to shield iris data at once. The eye's iris weave is a top ID mark, but stored iris data raise privacy fears. Their guess: the style part of the weave IDs folks and shrugs off turns and tilts.
Style traits beat stock traits by clear gaps in their ID tests. A style swap can then blur the ID part of an iris image. The blurred image still serves gaze and eye-part tasks well. The pair opens a path to safer eye ID systems, they say.
Catalog record
| CREATOR | Mengdi Wang, Efe Bozkir, Enkelejda Kasneci |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 52 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2503-04707
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Iris Style Transfer: Enhancing Iris Recognition with Style Features and Privacy Preservation through Neural Style Transfer Mengdi Wang, Efe Bozkir, Enkelejda Kasneci Abstract . Iris texture is widely regarded as a gold standard biometric modality for authentication and identification. The demand for robust iris recognition methods, coupled with growing security and privacy concerns regarding iris attacks, has escalated recently. Inspired by neural style transfer, an advanced technique that leverages neural networks to separate content and style features, we hypothesize that iris texture’s style features provide a reliable foundation for recognition and are more resilient to variations like rotation and perspective shifts than traditional approaches. Our experimental results support this hypothesis, showing a significantly higher classification accuracy compared to conventional features. Further, we propose using neural style transfer to obfuscate the identifiable iris style features, ensuring the protection of sensitive biometric information while maintaining the utility of eye images for tasks like eye segmentation and gaze estimation. This work opens new avenues for iris-oriented, secure, and privacy-aware biometric systems. Journal: PACMCGIT Volume: 8 2 21 6 DOI: 10.1145/3729413 CCS: Security and privacy Biometrics CCS: Security and privacy Privacy-preserving protocols CCS: Computing methodologies Image manipulation CCS: Human-centered computing † † cc-license: by 1. Introduction Reliable and robust automatic recognition of users using biometrics ( Jain et al., 2004 ; Jain, 2007 ) has been longed for decades. The iris, a rich source of unique biometric information, features distinctive textures and patterns that can reliably differentiate individuals. Due to its identifiability, uniqueness, and favorable mathematical properties ( Flom and Safir, 1987 ; Daugman, 2003 ; Daugman, 2009 ; Jain et al., 2004 ; Ma et al., 2004 ) , the iris has come to be regarded as a gold standard in biometrics. This has led to the widespread adoption of iris-based identification and authentication methods across a variety of devices, from dedicated iris scanners to VR/AR/MR (virtual/augmented/mixed reality) headsets and smart glasses, where reliability and convenience are paramount. Notably, Apple’s Vision Pro integrates Optic ID for iris-based authentication ( Apple, 2024 ) , and Microsoft’s HoloLens 2 also supports iris login on top of Windows Hello ( Microsoft, 2024 ) . While most head-mounted displays currently lack built-in iris recognition functionality, devices such as Meta Quest Pro and Magic Leap 2, which include eye-tracking capabilities, have promising potential to enable software-based iris recognition, as eye-tracking and iris recognition techniques share several underlying principles. Despite the expanding use of iris recognition, the impact of factors like image quality, sensor setup, and feature robustness under these varying conditions remains under-explored. At the same time, the popularity of iris recognition has sparked a rise in security threats like presentation attacks ( Sequeira et al., 2016 ; Kohli et al., 2017 ; Narkar and David-John, 2024 ) . Although countermeasures have been developed ( Sequeira et al., 2014b ; Sequeira et al., 2014a ; Doyle and Bowyer, 2015 ; Rigas and Komogortsev, 2015 ; Raja et al., 2015 ; Czajka and Bowyer, 2018 ; Raju et al., 2022 ) , concerns about the balance between privacy, security, and utility in iris-based systems continue to grow. Striking a sustainable trade-off between protection and usability ( John et al., 2020a ) is now more critical than ever. Neural style transfer ( Gatys, 2015 ; Gatys et al., 2016 ) opens up a fresh opportunity for tasks like image synthesis and artistic blending. It operates on the premise that each image is a combination of content and style. By leveraging convolutional neural networks (CNN), which are commonly trained for image classification tasks, neural style transfer strives to separate content and style features and then recombine them in a unique way. It provides a novel approach to interpret style features of an image statistically by capturing correlations between feature activations and can be hence regarded as a special domain adaptation problem ( Li et al., 2017 ) . As the style features of different images extracted by neural style transfer show quite distinctive properties, in this paper, we propose using iris style features for iris recognition. Considering that style features are extracted in the form of global statistical distribution, we hypothesize that iris style features, represented as global statistical distributions, are inherently robust to image variations such as rotation and perspective transformation due to their non-dependence on individual pixel positioning. In this regard, we investigate the impact of random rotation and perspective transformation of different degrees on iris recognition using traditional CNN features and style features on the OpenEDS2019 ( Garbin et al., 2019 ) dataset. Our results show that style features deliver significantly better test classification accuracy in both cases. Additionally, we suggest using neural style transfer to erase the sensitive iris style features to obscure the biometric identities of users without ruining image quality and the utility of eye images for downstream tasks like eye semantic segmentation and gaze estimation. We validate the effectiveness of our privacy-preservation method by presenting a prominent drop in recognition accuracy and a little to no decrease in segmentation accuracy on the OpenEDS2019 dataset, as well as a neglectable increase in gaze estimation error on the OpenEDS2020 dataset ( Palmero et al., 2020 ; Palmero et al., 2021 ) . The main contributions of this work are four-fold: • We propose a novel feature extraction method for the iris feature using neural style transfer, and suggest the use of the iris style feature for iris recognition accordingly. • We investigate the impact of random image rotation and perspective transformation of different transformation degrees, which are common problems with cameras as input modality, on iris recognition using conventional CNN features and style features. • We show that iris style features are remarkably more robust against random rotation and perspective transformation than traditional CNN features. Our results lay the foundation of iris recognition utilizing iris style features. • We suggest the use of neural style transfer to blend out the sensitive iris style features in eye images to conceal user identity from imposters without sacrificing much the quality and the utility of eye images. We empirically validate the feasibility of our approach with a significant drop in recognition accuracy and negligible to zero changes in image quality, eye segmentation accuracy, and gaze estimation error. Besides, we show that the risk of malicious use of our work by adversaries is relatively low. To the best of our knowledge, this is the first work to present neural style transfer for iris recognition and privacy preservation. The proposed iris style transfer prototype can be easily extended from static eye images to dynamic eye videos in real time on modern devices. Our work addresses a new research gap in this regard. We also publish our code 1 1 1 https://gitlab.lrz.de/hctl/Iris-Style-Transfer . for reproducibility. 2. Related Works 2.1. Iris Recognition As iris contains unique biometric features and demonstrates distinct mathematical advantages ( Flom and Safir, 1987 ; Daugman, 2003 ; Daugman, 2009 ; Jain et al., 2004 ; Ma et al., 2004 ) , it has been researched as user identification and authentication modality for long. The idea of iris recognition can be traced back to the early days when iris features were usually encoded by handcrafted filters and kernels (