초록/요약

Bio-modalities, such as the face, iris, and fingerprint, are ideal for establishing authentication in the futuristic networks, such as the medical cyber physical systems (MCPSs). In such a network, to authenticate and classify the bio-modalities, raw data would be traditionally sent to the cloud other than the proximal devices as they are resource-constrained. Thus, the centralized cloud-based solution not only incurs significant delay but also violates the data privacy as the data are moved to the cloud. In recent years, privacy-preserving on-device AI nodes are getting attention to solve certain classification problem, which can also be applied for classifying spoofed and real bio-modalities for authentication. To this end, we propose an on-device AI-based MCPS architecture, where the on-device AI node runs a light-weight but powerful classification algorithm, as we call it the feature-augmented random forest (FA-RF). The FA-RF combines the power of random forest with feature selection and a proposed feature augmentation mechanism. Besides privacy-preserving of the raw data, the proposed approach can significantly reduce the communication delay imposed on the network as cloud computation and communication is removed. Our proposal is verified on real datasets of the face, iris, and fingerprint bio-modalities provided by the Warsaw, Replay-Attack, and LiveDet 2015 Crossmatch benchmark, respectively. The experimental results show that our model can outperform the state-of-the-art architectures in four out of six tests. Besides, we show that the FA-RF can significantly reduce the training and testing time in both the cloud and the on-device AI node. © 2013 IEEE.