Producing secure multimodal biometric descriptors using artificial neural networks

With the growing importance and sensitivity of information and digital services being provided for different users, the need for better protection schemes has surfaces, according to the sensitivity of traditional secret-based techniques to simple attacks, such as shoulder surfing and guessing. Biometric authentication has emerged as a solution to these limitations, according to the high robustness of biometric features and their high availability. Combined with anti-spoofing methods that the biometric collection sensors are equipped with, biometric authentication has been able to provide better security to the protected systems. Furthermore, multimodal biometric authentication has been widely investigated in recent years to improve the accuracy of biometric authentication, by using multiple biometric templates to recognize a candidate. Despite the significant improvement of recognition accuracy, multimodal biometric authentication does not contribute towards securing the vulnerabilities of these systems. Several of these attacks occur during the transportation of the information over the network, which encouraged the proposal of methods that deny such attacks. However, existing methods require additional processing to detect these attacks, which increases the complexity of the system. Hence, a new method is proposed in this study to produce a fixed-size descriptor for each candidate based on their face and fingerprint templates. The proposed method also takes into consideration the timestamp in which the template is collected and a system identifier, so that, the descriptor becomes invalid after a predefined period of time and a descriptor generated at a certain system fails to authenticate of another, even if these descriptors are generated at the same timestamp. The proposed method uses a Convolutional Neural Network (CNN) to generated a 512-value descriptor, which is used to authenticate the candidate by simply measuring the Euclidean distance to the model descriptor. The embedding of the timestamp and system identifier in the descriptor allows the validation of these parameters in the same measurement, i.e. no additional processing is required. In addition to the high recognition accuracy achieved by the proposed method, with 99.41% and 99.32% accuracies using two different setups, the proposed method has also been able to maintain the privacy of the users, based on the one-way computations in the CNN.