An Empirical Study on Detecting Deception and Cybercrime Using Artificial Neural Networks

Ubiquity of the Internet and wide adoption of the computing and mobile devices is driving explosion of data. Interestingly, cybercriminals are also leveraging these popular technologies to cash in on cybercrime in form of scams, fraud and fake online reviews. Existing content filtering techniques, which have been successful in containing spam, are failing to filter these new types of cybercrime because cybercriminals generate text messages to bypass content filters. In this paper, we use natural language processing and a deception-detection discourse to build hybrid models for detecting these forms of text-based cybercrime. Since we have four datasets each of which contains deceptive text messages representing a specific type of cybercrime and truthful text messages, we combine 2 datasets and 3 datasets together to generate training sets for the hybrid models with more than one type of cybercrime. The hybrid cybercrime detection models are trained using Artificial Neural Networks (ANN), Naive Bayes (NB), Support Vector Machines (SVM) and kth Nearest Neighbor (kNN). The models are then evaluated on test sets containing instances that were not part of the training sets. The results for model performance of NB, kNN and SVM classifiers are compared against those of ANN. Most the models generalize well in detecting cybercrime. ANN model performance on the test sets ranges from 70% to 90% accuracy compared to model performance range of 60% to 80% for the other three classifiers. The best performance is in detecting unfavorable fake reviews and fraud.