Explainable AI for Classifying Devices on the Internet
Artūrs Lavrenovs, Roman Graf · 2021
Devices reachable on the Internet pose varying levels of risk to their owners and the wider public, depending on their role and functionality, which can be considered their class. Discussing the security implications of these devices without knowing their classes is impractical. There are multiple AI methods to solve the challenge of classifying devices. Since the number of significant features in device HTTP response was determined to be low in the existing word-embedding neural network, we elected to employ an alternative method of Naive Bayes classification. The Naive Bayes method demonstrated high accuracy, but we recognise the need to explain classification results to improve classification accuracy. The black-box implementation of Artificial Neural Networks has been a serious concern when evaluating the classification results produced in most fields. While devices on the Internet have historically been classified manually or using trivial fingerprinting to match major vendors, these are not feasible anymore because of an ever-increasing variety of devices on the Internet. In the last few years, device classification using Neural Networks has emerged as a new research direction. These research results often claim high accuracy through the validation employed, but through random sampling there always occur devices that cannot be easily classified, that an expert intuitively would classify differently. Addressing this issue is critical for establishing trust in classification results and can be achieved by employing explainable AI. To better understand the models for classifying devices reachable on the Internet and to improve classification accuracy, we developed a novel explainable AI method, which returns the features that are most significant for classification decisions. We employed a Local Interpretable Model-Agnostic Explanations (LIME) framework toexplain Naive Bayes model classification results, and using this method were able to further improve accuracy with a better understanding of the results.