Towards Feasible Continuous Authentication Using PPG Signal With Deep Autoencoders

This Thesis proposes a session-based continuous authentication scheme using photoplethysmography (PPG). The proposed authentication model utilizes the ability of deep autoen-coders to extract relevant features from the row PPG signal. Then, the model uses an outlier detection algorithm (namely Local Outlier Factor) to identify adversaries’ presence in the system. The LOF algorithm needs only to be trained on the honest user data, making it suitable to apply in practical use-cases. Furthermore, The proposed model was tested on the CapnoBase, and BIMDC benchmarking dataset and achieved an F1 score of 89.45% and 93.32%, respectively. The proposed solution was compared to state-of-the-art and achieved a 15% lower EER in relative terms on the bidmc dataset and a higher F1 score by more than 4% on the CapnoBase dataset. Furthermore, the proposed solution provides a faster run time than the competing state-of-the-art solution by requiring around 0.5 seconds for processing 18,000 beats, compared to 3.6 seconds for state-of-the-art —an ≈ 85% reduction.