R. Donida Labati, A. Genovese, V. Piuri, F. Scotti, S. Vishwakarma, I-SOCIAL-DB: A labeled database of images collected from websites and social media for iris recognition, Image and Vision Computing, vol. 105, no. 104058, pp. 1-9, January 2021, ISSN 0262-8856.
Biometric systems
Iris Deidentification
Iris Deidentification investigates methods for protecting biometric privacy in high-resolution facial photographs shared on websites, social networks and other public platforms. Improvements in consumer cameras and iris-recognition algorithms have made it possible to extract useful ocular information even from ordinary portrait images acquired in uncontrolled conditions. Since iris texture is highly distinctive and remains stable over time, publicly available photographs may potentially be used to identify, recognise or track individuals without their knowledge or explicit consent.
The project developed an automatic deidentification approach based on generative adversarial networks. Rather than obscuring the eyes through blurring, pixelation or masking, the proposed method generates a visually realistic replacement for the iris region, removing and substituting the biometric information associated with the original individual while maintaining the natural appearance of the face. This makes the protected image suitable for publication and visual interpretation, while substantially reducing the probability that iris-recognition systems can associate it with the subject’s true identity. Experimental evaluations considered both biometric privacy and image realism, showing that the generated samples can significantly reduce recognition risks and, in many cases, remain difficult to distinguish from genuine images.
The research was supported by I-SOCIAL-DB, a labelled database created from publicly available websites and social media. The dataset contains 3,286 ocular regions extracted from 1,643 high-resolution facial images of 400 individuals, with manually annotated iris boundaries, occlusions and reflections. Together with dedicated recognition protocols and performance measures, this resource enables the systematic study of iris recognition in unconstrained images and the quantitative assessment of privacy-preserving transformations. The project therefore contributes a practical framework for balancing personal-data protection, visual quality and the legitimate use of online imagery.
The project developed an automatic deidentification approach based on generative adversarial networks. Rather than obscuring the eyes through blurring, pixelation or masking, the proposed method generates a visually realistic replacement for the iris region, removing and substituting the biometric information associated with the original individual while maintaining the natural appearance of the face. This makes the protected image suitable for publication and visual interpretation, while substantially reducing the probability that iris-recognition systems can associate it with the subject’s true identity. Experimental evaluations considered both biometric privacy and image realism, showing that the generated samples can significantly reduce recognition risks and, in many cases, remain difficult to distinguish from genuine images.
The research was supported by I-SOCIAL-DB, a labelled database created from publicly available websites and social media. The dataset contains 3,286 ocular regions extracted from 1,643 high-resolution facial images of 400 individuals, with manually annotated iris boundaries, occlusions and reflections. Together with dedicated recognition protocols and performance measures, this resource enables the systematic study of iris recognition in unconstrained images and the quantitative assessment of privacy-preserving transformations. The project therefore contributes a practical framework for balancing personal-data protection, visual quality and the legitimate use of online imagery.