@MASTERSTHESIS\{IMM2016-07053,
    author       = "M. K. Sigaard",
    title        = "Protecting Templates for Multimodal Biometrics on Smartphone - Using Binarized Statistical Image Features with Bloom Filters",
    year         = "2016",
    school       = "Technical University of Denmark, Department of Applied Mathematics and Computer Science",
    address      = "Richard Petersens Plads, Building 324, {DK-}2800 Kgs. Lyngby, Denmark, compute@compute.dtu.dk",
    type         = "",
    note         = "{DTU} supervisor: Christian D. Jensen, cdje@dtu.dk, {DTU} Compute",
    url          = "http://www.compute.dtu.dk/english",
    abstract     = "Every day billions of people use their smartphones to access online services. Companies such as Apple, Google, and Samsung have made it possible to use fingerprint for authentication. Recently the companies launched their own payment service, called Apple Pay, Android Pay, and Samsung Pay, with the purpose of making it easier to shop online. It was made possible to access the service using fingerprint. One of the biggest issues with biometric authentication is that if a fingerprint is stolen, then it cannot be used for biometric authentication again, as it would be too risky. There exist a number of different attacks against biometric systems but one of the most damaging is against the database containing the biometric templates. The goal of this thesis is to use Bloom filters to obtain biometric template protection for a biometric authentication system on smartphones. Face and periocular regions are the biometric characteristics that are used to perform authentication in this thesis. Features from the biometric characteristics are extracted by using Binarized Statistical Image Features (BSIF) that describes the texture of an image by folding it with {BSIF} filters and thus describe each pixel by its nearest neighbouring pixels. Biometric template protection is defined in the {ISO} standard as when a stored biometric template is unlinkable and irreversible. Unlinkable means that it is not possible to cross reference a user’s stored biometric template and determine if the user is present in two different databases. Irreversible means that it is not possible or at least computationally very difficult to reverse engineer a protected biometric template and get the original biometric template. A Bloom filter is a space-efficient, randomized data structure that selects members with a probability of false positives but not false negatives. The biometric templates will be converted to Bloom filters and will thus become irreversible, while still being usable for authentication. The created Bloom filters are then permuted by a key that is unique for each user, to achieve that the created Bloom filter templates are unlinkable between different templates. The system is created in an Android environment and tested on a Samsung S5 smartphone. The feature extraction and Bloom filter creating are implemented in a C++ library. This is done in order to run additional tests on a database containing face images captured with a Samsung S5 phone’s camera, and this also makes it easier to use the created system in other smartphone environments such as {IOS}. For the images in the database facial recognition had the best accuracy with a {GMR} at 82.68 \% at a {FMR} at 0.01 \% with Bloom filters, and a {GMR} at 90.05 \% at a {FMR} at 0.01 \% without Bloom filters. Different kinds of comparison score and feature fusion were used in order to improve the accuracy of the system. Comparison score fusion had the best performance for the database images. With Bloom filters the {GMR} was 91.47 \% at {GMR} 0.01 \% and without Bloom filters the {GMR} was 96.05 \% at a {FMR} at 0.01 \%. The proposed system was tested by 22 subjects in a practical scenario on Samsung Galaxy S5 phone. The test indicated the applicability of the template protected authentication system, in that the accuracy performance of the template protected system was decreased but similar to the same system without template protection."
}