CRII: SaTC: Automatic Generation of API to Natural Language Data Type Mappings for Developer and End User Privacy Risk Mitigation

自从智能手机出现以来，越来越多的人可以使用可上网的软件应用程序（app）。这一点，再加上移动设备上的各种传感器，使公众极易受到隐私风险的影响，因为敏感信息（例如位置、相机图像、生物特征）可能会泄露到互联网上。为了帮助用户在使用应用程序时对潜在的隐私风险做出明智的决定，监管机构越来越多地要求应用程序开发人员包括隐私政策，说明收集或共享的信息以及如何使用这些信息。然而，即使存在这样的隐私政策，也必须信任应用程序开发人员遵守其中的承诺。此外，开发人员对他们的策略的遵守负责，并且必须确信他们的隐私策略准确地代表了他们的实践。该项目旨在帮助开发者和一般应用程序用户验证隐私政策和他们所代表的应用程序的一致性，通过生成一个自动化的过程，将隐私政策中使用的语言语义与用于生成应用程序本身的代码联系起来。此外，项目将使用这个框架为最终用户和开发人员生成工具，从而直接从这项工作中受益。该研究项目旨在使用机器学习生成代码级api和自然语言数据类型之间映射的自动化过程。生成的映射将在开发人员和最终用户工具中使用，以识别和帮助减轻开发和应用程序使用过程中潜在的隐私泄露。隐私策略和应用程序代码之间的不一致检测的当前状态需要手动生成从代码级应用程序接口（API）方法到面向隐私的自然语言数据类型的映射。即使是小型应用类别，这个过程也可能需要人工检查数千种方法和数百个注释，这可能会导致由于疲劳和不完整的领域知识而导致不准确。api也会随着方法的引入而改变，而弃用会导致过时的映射。随着环境的不断发展，这些问题使得实际应用框架变得困难。本项目将通过两方面的贡献来应对这些挑战。首先，机器学习将应用于映射生成过程，以生成一种自动化的、可扩展的方法，用于根据需要为api生成代码短语映射。这将允许对API级别、方法和应用程序类别进行比对检测，而不仅仅是在以前的贡献中构建的。这种自动化方法将利用最先进的预训练语言模型来检测API文档和隐私策略中使用的自然语言数据类型之间的语义相似性。其次，来自自动化模型的结果映射将应用于实际的开发人员和最终用户工具，以便为减轻隐私风险做出明智的决策。polidoid工具套件将包括一个面向开发人员的集成开发环境插件，该插件可以根据隐私策略检测潜在的意外隐私泄露，并为最终用户提供实时错位检测工具。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Since the advent of the smart phone, an increasing amount of the population has gained access to Internet-accessible software applications (apps). This, coupled with the various sensors available on mobile devices, make the general public highly susceptible to privacy risks as sensitive information (e.g., location, camera images, biometrics) may be leaked to the Internet. To help users make informed decisions about the potential privacy risks in using apps, regulators increasingly require app developers to include privacy policies communicating what information is collected or shared and how that information is used. However, even when such privacy policies are present, trust must be put in the app developers to adhere to the promises therein. Furthermore, developers are accountable for their adherence to their policies and must be confident that their privacy policies accurately represent their practices. This project aims to assist both developers and general app users in verifying the alignment of privacy policies and the apps they represent by producing an automated process for linking the semantics of language used in privacy policies with the code used to produce the apps themselves. Furthermore, the project will use this framework to generate tools for end users and developers to directly benefit from this work.The research project aims to produce an automated process for generating mappings between code-level APIs and natural language data types using machine learning. The resulting mappings will be utilized in developer and end user tools to identify and help mitigate potential privacy leakage during development and app usage. The current state of misalignment detection between privacy policies and app code requires the manual generation of mappings from code-level Application Program Interface (API) methods to privacy-oriented natural language data types. Even for small app categories, this process can require a human to review thousands of methods and hundreds of annotations resulting in potential for inaccuracies due to fatigue and incomplete domain knowledge. APIs also change as methods are introduced and deprecated resulting in outdated mappings. These problems make it difficult to apply the framework practically as the environment continually evolves. This project will address these challenges through two contributions. First, machine learning will be applied to the mapping generation process to produce an automated, scalable method for generating code-phrase mappings for APIs as needed. This will allow for misalignment detection for API levels, methods, and app categories beyond those build in previous contributions. This automated approach will make use of a state-of-the-art pre-trained language models to detect semantic similarity between API documentation and natural language data types used in privacy policies. Second, the resulting mappings from the automated model will be applied to practical developer and end user tools to enable informed decision for privacy risk mitigation. The PoliDroid tool suite will be developed including a developer-oriented integrated developer environment plugin which detects potential unintended privacy leaks based on a privacy policy and a real-time misalignment detection tool for end users.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.