Evaluation of Unsupervised Deep Learning Methods for Detecting Malicious Activities in Mobile Browsing Data

This study explores the detection of malicious activity in mobile browsing histories using three unsupervised deep learning algorithms: autoencoders, variational autoencoders and generative adversarial networks. Data are collected from simulated scenarios combining legitimate and malicious activities, then pre-processed to capture temporal sequences. Evaluation is based on precision, recall and F1 score, as well as computational efficiency. The results show that AEs effectively detect simple anomalies, while VAEs identify subtle patterns in complex distributions. GANs, by generating synthetic data, improve model robustness. This study highlights the advantages and limitations of each approach, and proposes hybrid methods for enhancing threat detection in mobile environments.