MemberGuard: Protecting Machine Learning Privacy from Membership Inference

机器学习已经成为许多现实世界应用的核心部分。然而，机器学习模型容易受到成员推理攻击。在这些攻击中，攻击者可以推断给定的数据记录是否属于模型的训练数据的一部分。在这个项目中，该团队的目标是开发可用于对抗这些攻击的新技术，例如1)新的成员泄漏分析模型，2)易感性诊断的新方法，3)利用隐私和效用的新防御。据估计，面向数据的服务在未来将是有价值的资产。这些技术可以帮助澳大利亚在机器学习安全和隐私方面获得领先优势，并保护其在这些服务上的知识产权。

Machine Learning has become a core part of many real-world applications. However, machine learning models are vulnerable to membership inference attacks. In these attacks, an adversary can infer if a given data record has been part of the model's training data. In this project, the team aims to develop new techniques that can be used to counter these attacks, such as 1) new analytical models for membership leakage, 2) new methods for susceptibility diagnosis, 3) new defences that leverage privacy and utility. Data-oriented services are estimated to be valuable assets in the future. These techniques can help Australia gain cutting edge advantage in machine learning security and privacy and protect its intellectual property on these services.