Automatic evaluation of privacy policy:a machine learning approach

Privacy is taking an increasingly prominent place in today’s digital world. People wish to control their private information when interacting with websites on the Internet. However, customer information is one of the keys to successful Internet business. Internet companies wish to gather and use as much customer information as possible in order to facilitate business, build competitive barrier and generate profits. In the middle of the conflicting notions, privacy policy serves as the main channel for the companies to disclose their practice for dealing with user’s private information. However, many surveys have shown that users seldom read the privacy policies and the current mechanisms to present website privacy policies have not been successful. The readability issue of privacy policies calls for automated ways of privacy policy evaluation to assist users to quickly gain insights about the privacy practice of the website. This research addresses the present gap in the communication and understanding of privacy policies, by creating an automated privacy policy evaluation framework that provides automatic categorization, analysis and grading of privacy policies. We advocate a machine learning approach towards privacy policy evaluation and lay the fundamental basis for this new approach. We present a privacy policy evaluation framework. The framework comprises several core components, such as privacy policy paragraph categorization, privacy policy grading, share statement understanding, as well as some accessorial components, such as privacy policy detection, search result extraction, text extraction, visualization, and web application interface. We define the common categories of privacy policies, label real-world privacy policies to form the datasets and implement all the aforementioned components. We investigate the application of text classification to categorize privacy policy paragraphs. We present extensions to this categorization scheme, such as twolayered classification, multi-label classification, grading and visualization. We propose two approaches for the task of share statement understanding, each with different variants. We extensively experiment with our proposed schemes and approaches on the privacy policy datasets. Our study results demonstrate that the machine learning approach is effective in building privacy policy evaluation systems that serve the purpose to assist users to better understand the privacy policies.