Deep Learning Based Behavior Anomaly Detection within the Context of Electronic Commerce

To address the threat of online misbehaviors by authenticated e-commerce (EC) consumers in exploiting business logic and regulation loopholes, this paper proposed Event-Based Behavior Analysis (EBBA), a reconfigurable framework with multiple deep-learning (DL) time-series forecasting models to detect and examine online behavioral anomaly. A prototype EBBA with two switchable DL models, Recurrent Neural Networks (RNN) and Long/Short Term Memory (LSTM), was constructed in Python and trained by real-world access logs from a tier-one EC company in Taiwan. Preliminary experimental results show that RNN and LSTM models achieved as high as 86.4% and 92.2% correctness ratios, with as low as zero and 0.0385% mis-granting ratios in prediction of online misbehavior occurrences, respectively.