"Security Mobile App User Reviews Classification"

"Mobile app user reviews contain valuable feedback about software functionality, quality, and security. Existing approaches for analyzing security-related reviews often rely on traditional feature extraction methods, limiting their ability to detect nuanced security concerns. This study aims to develop an automated framework for classifying mobile app user reviews into security-related and non-security-related categories. The framework leverages semantic representations and an iterative active learning process to improve annotation efficiency and model performance, while a web-based graphical user interface (GUI) with an integrated large language model (LLM) provides interpretable explanations, supporting both software engineers and end-users in evaluating mobile app security. We collected over one million reviews from multiple sources and applied a three-stage annotation pipeline: keyword-based filtering, expert validation, and iterative active learning that combines sentence-transformer embeddings with multiple machine learning classifiers. Using the resulting labeled dataset, two transformer-based models (SBERT and Paraphrase) were fine-tuned for binary classification, with the GUI+LLM providing explanations for predicted labels.  On 10-fold cross-validation, SBERT achieved 88.57\\% accuracy and 88.65\\% F1-score, while Paraphrase reached 90.38\\% accuracy and 90.40\\% F1-score. On the held-out test set, SBERT obtained 74.11\\% accuracy and 73.94\\% F1-score, whereas Paraphrase achieved 88.39\\% accuracy and 88.38\\% F1-score. Transformer-based models with sentence embeddings can effectively classify security-related user reviews, with Paraphrase outperforming SBERT. The integrated GUI and LLM enhance interpretability, providing a practical tool for supporting security assessment in mobile applications."