Introduction
Modern smartphone operating systems, particularly Android 6.0 and later, utilize an "ask-on-first-use" (AOFU) policy to manage sensitive permissions. While this approach enhances transparency by prompting users at runtime, it introduces a significant cognitive burden. Most users lack the technical expertise to determine if a specific permission request—such as a calculator app asking for location data—is a functional necessity or a privacy overreach. Existing recommendation systems often rely on historical user decisions, which can inadvertently reinforce poor privacy habits. To address these challenges, we present AutoPer+, an autonomous system designed to provide intelligent "Allow," "Deny," or "Ask" recommendations.
Methodology
AutoPer+ operates on the principle that an app's natural language description serves as a functional contract. By analyzing what an app claims to do and comparing it to the behavior of similar apps, the system can identify anomalous permission requests. The framework consists of three primary modules:
- Multi-topic Permission Mapping: Using Latent Dirichlet Allocation (LDA), the system extracts functional topics from app descriptions. It then employs Mutual Information (MI) to rank permissions based on their relevance to these topics.
- Similar-App Mining: To establish a baseline for "normal" behavior, AutoPer+ identifies functionally similar apps. It utilizes Word2Vec for text embedding and a Bidirectional LSTM (BiLSTM) with an attention mechanism to classify apps into 118 fine-grained categories. A semi-supervised learning approach is used to leverage a large set of unlabeled apps alongside a smaller set of manually labeled representative apps.
- Decision & Explanation Engine: This module reconciles topic-based relevance with peer-usage patterns. If a permission is standard for the app's niche, it recommends "Allow." If it is anomalous, it recommends "Deny." In cases of low confidence, it recommends "Ask" and provides a detailed explanation to the user.
Experimental Evaluation
The system was evaluated using a comprehensive dataset of 31,023 Android apps from the Google Play Store. The evaluation focused on 24 "Dangerous" permission groups.
- Accuracy: AutoPer+ achieved an overall recommendation accuracy of 84.1%.
- Classification Performance: The BiLSTM model achieved an F1-score of 0.864 in similar-app classification, significantly outperforming traditional machine learning baselines like Random Forest.
- Permission-Specific Success: The system was particularly effective at identifying unnecessary permissions for sensitive categories, such as CALENDAR (98.12% F1-score) and LOCATION (92.89% F1-score).
A user study involving 37 participants confirmed the system's practical utility. Users reported high satisfaction with the "Ask" feature, which provided transparency and allowed them to audit their previous privacy decisions.
Conclusion
AutoPer+ demonstrates that natural language descriptions and peer-behavior analysis are powerful tools for automating privacy management. By moving beyond simple historical tracking, the system provides a proactive defense against permission overreach. Future work will focus on integrating dynamic analysis to detect deceptive descriptions and incorporating contextual factors, such as location or time of day, into the recommendation engine.
Summary updated