Automated Authentication of Large-Scale IoT Devices with Hybrid Feature Selection

Abstract

Nowadays, large-scale IoT (Internet of Things) devices are connected to the network. However, due to the simplicity, limited memory and poor computing power of most IoT devices, it is difficult to apply complex security authentication protocols to these devices. Many attackers use this loophole, impersonating legitimate devices to launch malicious attacks on the network, threatening the security of other legitimate devices. Therefore, in order to strengthen the identification and management of devices, the existing device fingerprint identification technology based on machine learning is to extract device traffic at the access gateway and generate device fingerprinting to realize the identification of IoT devices. However, to achieve high recognition accuracy, existing device fingerprint algorithms need to extract a large number of fingerprint features from network traffic for machine learning, which increases the complexity of implementation and delay of device identification. Also, it is not suitable for the current large-scale access environment. Therefore, this paper proposes a hybrid feature selection method to preprocess features for the optimal feature subset, which can not only reduce the feature dimension, but also improve the accuracy of device recognition. Firstly, the method filters the irrelevant features according to the characteristic fluctuation and the entropy of correlation information between the feature and the target. Then, based on the classification accuracy of the learning algorithm, a random search method—forest optimization algorithm is used to further optimize the filtered feature subsets under the evaluation criteria. Experiments show that this method can reduce feature dimension by 81.0% and improve classification accuracy from 86.4% to 93.1%.