PrivInfer - Programming Languages for Differential Privacy: Conditioning and Inference

每天都有大量的个人数据被收集。这些数据可能对科学和医学研究或目标业务非常有价值。不幸的是，隐私问题限制了这些大量信息的使用和发布。已经提出了几种技术，目的是使数据匿名。然而，当攻击者可以利用额外的知识时，这些技术就失去了它们的有效性。差分隐私是一种有前途的隐私保护数据发布方法：它提供了一个强有力的保证，限制了用户I由于参与差异隐私数据分析而遭受的伤害的增加，即使在最坏的情况下也是如此。确保差异隐私的标准方法是在数据分析结果中添加一些统计噪音。差分私有机制已经被广泛地应用于统计分析、组合优化、机器学习、分布式计算等领域。此外，一些编程语言验证工具也被提出来帮助程序员验证给定程序是否是差分私有的，这些工具在验证使用标准机制的差分私有程序方面已经被证明是成功的。然而，他们只提供有限的支持推理差异隐私时，这是使用非标准的机制.一个局限性来自于这些工具内置的简化概率模型。特别是，这些简化的模型不提供支持（或只有非常有限的支持）推理显式条件分布和概率推理。从验证的角度来看，处理显式条件分布是困难的，因为它需要在验证工具的内部逻辑中找到事件和概率度量的可管理表示。此外，它需要一组原语来有效地处理它们。在这个项目中，我们的目标是通过扩展差分隐私验证工具的范围来克服这些限制，以避免关于条件分布和概率推理的显式推理。对条件分布和概率推理的支持对于机器学习算法的推理至关重要。这些都是实现高效和准确的数据分析的重要工具，以收集大量的数据。因此，该项目的目标是提供一种新的编程语言技术，用于增强基于机器学习的隐私保护数据分析。

An enormous amount of individuals' data is collected every day. Thesedata could potentially be very valuable for scientific and medicalresearch or for targeting business. Unfortunately, privacy concernsrestrict the way this huge amount of information can be used andreleased. Several techniques have been proposed with the aim ofmaking the data anonymous. These techniques however lose theireffectiveness when attackers can exploit additional knowledge.Differential privacy is a promising approach to the privacy-preservingrelease of data: it offers a strong guaranteed bound on the increasein harm that a user I incurs as a result of participating in adifferentially private data analysis, even under worst-caseassumptions.A standard way to ensure differential privacy is by adding somestatistical noise to the result of a data analysis. Differentiallyprivate mechanisms have been proposed for a wide range of interestingproblems like statistical analysis, combinatorial optimization,machine learning, distributed computations, etc. Moreover, severalprogramming language verification tools have been proposed with thegoal of assisting a programmer in checking whether a given program isdifferentially private or not.These tools have been proved successful in checking differentiallyprivate programs that uses standard mechanisms. They offer however only alimited support for reasoning about differential privacy when this isobtained using non-standard mechanisms. One limitation comes from thesimplified probabilistic models that are built-in to those tools. Inparticular, these simplified models provide no support (or only verylimited support) for reasoning about explicit conditionaldistributions and probabilistic inference. From the verificationpoint of view, dealing with explicit conditional distributions isdifficult because it requires finding a manageable representation, inthe internal logic of the verification tool, of events and probabilitymeasures. Moreover, it requires a set of primitives to handle themefficiently.In this project we aim at overcoming these limitations by extendingthe scope of verification tools for differential privacy to supportexplicit reasoning about conditional distributions and probabilisticinference. Support for conditional distributions and probabilisticinference is crucial for reasoning about machine learningalgorithms. Those are essential tools for achieving efficient andaccurate data analysis for massive collection of data. So, the goal ofthe project is to provide a novel programming language technologyuseful for enhancing privacy-preserving data analysis based on machine learning.