TL;DR
This paper introduces a novel system, APPE, to analyze privacy policy adherence across over two million apps in the Apple iOS App Store. It reveals that a significant portion of apps either lack privacy policy links or provide broken/irrelevant links, challenging the perception of Apple as a privacy-centric platform.
What Problem Does This Paper Address?
Mobile apps collect vast amounts of user data, often without users fully understanding the implications. Regulatory frameworks like GDPR and FTC guidelines mandate that app developers provide clear privacy policies. However, previous research has shown that many apps, particularly on Google Play, fail to provide these policies or offer policies that are difficult to understand or misleading. This study aims to comprehensively assess the state of privacy policy adherence specifically within the Apple iOS App Store, a platform often perceived as more privacy-conscious, and to quantify the extent of compliance issues.
Methods
The core of this research is the App Privacy Policy Extractor (APPE), a system designed to automate the analysis of privacy policies at scale. APPE comprises three main components:
- App Store Crawler: This component systematically navigates the iOS App Store, categorizing apps and extracting metadata, including alleged privacy policy links. It processed over two million apps across 24 categories.
- Policy Scrapper: This component follows the extracted privacy policy links. It uses a headless Chrome browser to capture dynamic content and supports various document formats (HTML, PDF, DOC, DOCX, TXT). It also checks for domain validity, HTTP response codes (identifying errors like 404 Not Found, 403 Forbidden), and determines the language of the document.
- Privacy Policy Recogniser (PPR): This is a deep learning model trained to classify whether a document is a valid English-language privacy policy. It utilizes a combination of LSTM and CNN architectures, leveraging transfer learning with pre-trained word embeddings (GloVe). The PPR classifies documents into "Privacy Policy" or "Non-policy."
Datasets
The study utilized a dataset of over two million apps from the Apple iOS App Store.
A new benchmark dataset, the App Privacy Policy Corpus (APPC-451K), was created. This corpus consists of over 451,000 verified English privacy policy documents, making it the largest such corpus available. It was constructed from the verified privacy policy links identified by the APPE system.
Key Insights and Ideas
- Apple's Privacy Stance Questioned: The findings challenge the notion that Apple's App Store is significantly more privacy-compliant than Google Play, suggesting a systemic issue across major app marketplaces.
- Widespread Non-Compliance: A substantial portion of apps fail to provide even basic privacy policy links, indicating a significant gap in developer responsibility.
- "Broken" Policies are Common: Even when links are provided, a large percentage lead to errors or irrelevant content, suggesting a lack of genuine effort to inform users.
- Language Barrier: A significant number of non-English privacy policies were found, highlighting the need for better enforcement of language requirements or translation tools.
- Category-Specific Issues: Certain app categories, like Entertainment and Games, show particularly low rates of privacy policy provision, despite having a large user base.
Key Results
- Link Availability: Only 58.5% of apps provided links to privacy policies. Conversely, 39.3% did not provide any links.
- Link Validity: Of the links provided, only 38.4% actually led to valid privacy policies. The remaining 61.6% were either errors (26.2%), non-policy documents (16.9%), or non-English documents (18.5%).
- Common Errors: The most frequent errors encountered were "404 Not Found" (124,533 occurrences) and "DNS Lookup" (113,081 occurrences).
- Non-English Policies: Spanish was the most prevalent non-English language for privacy policies, followed by German and Portuguese.
- APPC-451K Corpus: The study created a corpus of over 451,000 verified English privacy policies, with an average length of 1845 words.
Limitations and Future Work
- Language Focus: The study primarily focused on English privacy policies, acknowledging that non-English policies might still be valid but were excluded for analysis.
- Category Comparison Nuances: While comparisons between categories were made, the paper notes that the significant difference in the number of apps evaluated between this study and previous ones (e.g., Google Play studies) requires careful consideration when drawing direct comparisons.
- Future Work: The authors intend to conduct a more comprehensive analysis of the APPC-451K corpus. They also suggest that app stores should implement automated tools to verify the legitimacy of privacy policy links before submission. Further research could explore the readability and content of privacy policies in more detail.
Summary updated