An empirical study of supervised email classification in Internet of Things: Practical performance and key influencing factors

Abstract

Internet of Things (IoT) is gradually adopted by many organizations to facilitate the information collection and sharing. In an organization, an IoT node usually can receive and send an email for event notification and reminder. However, unwanted and malicious emails are a big security challenge to IoT systems. For example, attackers may intrude a network by sending emails with phishing links. To mitigate this issue, email classification is an important solution with the aim of distinguishing legitimate and spam emails. Artificial intelligence especially machine learning is a major tool for helping detect malicious emails, but the performance might be fluctuant according to specific datasets. The previous research figured out that supervised learning could be acceptable in practice, and that practical evaluation and users' feedback are important. Motivated by these observations, we conduct an empirical study to validate the performance of common learning algorithms under three different environments for email classification. With over 900 users, our study results validate prior observations and indicate that LibSVM and SMO-SVM can achieve better performance than other selected algorithms.

PF‐ITS: Intelligent traffic service recommendation based on DeepAFM model

“…40 To capture the relationships in features, the Wide & Deep model 41 combines the memory ability of the wide part with the generalization ability of the deep part in joint training, so as to win outstanding accuracy and scalability in most recommendation scenarios. 42 To overcome the low efficiency in artificial feature engineering, Guo et al 43 proposed DeepFM, in which the power of factorization machine (FM) and DNN are combined to reach significantly enhanced performance in effectiveness and efficiency. With pairwise feature combinations learned in FM 44 and high-order feature interactions captured in DNN, the DeepFM wins the Wide & Deep based on its merely raw features without feature engineering.…”

Feature autoencoder for detecting adversarial examples

“…The ubiquitous deep neural network (DNN) brings us to the era of artificial intelligence, ranging from speech recognition to identity verification. 1 More specifically, speech recognition technologies 2 permit machines to understand human voice, while face recognition technologies 3 enable machines to recognize a person according to his/her peculiarities, that is, fingerprints, face images, [4][5][6][7][8][9] and so on. However, DNNs are vulnerable to adversarial example attacks, [10][11][12][13] where an adversary attaches imperceptible perturbations to the clean image for inducing image-based model misclassification.…”

VISEL: A visual and magnetic fusion‐based large‐scale indoor localization system with improved high‐precision semantic maps

“…It is important to obtain an accurate location to enhance high-quality indoor location-based service, 1 including providing a safeguard for network access security 2,3 and behavior analysis of popular Internet of Things smart terminals. [4][5][6][7] Meanwhile, lower deployment costs promote the pervasive adoption of indoor location services.…”