CVPR Daily - Wednesday

about some genetic component. That is where the difference in gender is. The observation that we have been working on is that females usually have finer skin so the ridges have high density. We have more ridges. This data counts. There are some pictures that we constructed in order to distinguish males from females: we did find a difference. So far, we didn’t apply deep learning yet, but there are autoencoders that seem to work well with this constraint… so high accuracy matching, yet on the other side, we remove the soft biometric aspect. This can be for gender and age as well. There is a difference in matching for age as well because, as I mentioned, older people may have more issues in interacting with the sensors. Now, the future research on this part is we want to have data collection on mobile applications, but mainly on touch glass. So we need to study the quality of these images on touch glass sensors and interoperability. That’s why I’m still in this modality because I want to continue this part. Another aspect that I really like is the security of these systems. It’s been more than ten years that the security for these systems has been a big issue in the community. Mainly, the countermeasures against spoof attacks where fingerprints are replicated artificially with inexpensive materials like gelatin, latex, and playdough. These artifacts can actually carry the identity of someone and deceive the systems. These are cases like where a Japanese female could deceive the system at the airport. They can even alter the fingerprint. We really focus on detecting characteristics of vitality in fingerprints. The current detectors are biased based on the replica, which is the material used for making these spoofs. Let’s say in known attacks, all of them are in silicon, in gelatin: then the algorithms are trained only on those specific spoofs. So when we are testing the difference between different materials, it will probably fail. Another bias that we are working on is related to the spoofing method. The fingerprint that can be used for making an artifact can be obtained with the cooperation of the user, but also through non- cooperative methods. Let’s say we know that somebody is leaving fingerprints, since we do leave them on 14 Wednesday Emanuela Marasco