Iris liveliness detection using transfer learning on small cyber dataset

Iris recognition is one of the most popular and secure physical biometric identification methods, that have been realized to date. It is largely used in several applications nowadays, therefore, iris presentation attack detection (PAD) is of prime importance. This research features liveliness detection in which spoof (paper-printed) iris images are distinguished from real ones. The key challenge in this research was the utilization of small dataset in training a deep convolution neural network (CNN) to achieve maximum accuracy. Although deep CNNs are immensely popular and successful in solving computer vision and pattern recognition problems but, with small datasets it is highly probable to overfit data. To solve this problem, this research makes use of a popular machine learning technique, transfer learning, for iris liveliness detection. It utilized MobileNets architecture [1] pre-trained for object detection, reused it for iris liveliness detection by manipulating the model to properly identify specific biometrics. It was achieved by retraining the last layer with small iris dataset. Necessary changes in the model's hyperparameters and feature vectors were made for iris liveliness detection. High accuracies were achieved. This technique will provide a gateway into security features that are being adopted in modern technology and prevention of attacks improving cybersecurity, even with a limited amount of resources (data) to begin with