Towards understanding the interactions between the trustworthiness properties of machine learning

随着机器学习模型在我们的日常生活中变得无处不在，对相关伦理问题的担忧变得越来越重要。因此，我们正在见证许多公共和私人倡议的出现，呼吁人工智能在伦理上保持一致的发展。机器学习模型有望表现出几个可信的性质，例如公平、对隐私攻击(即成员关系推理、属性推理、模型提取)和安全漏洞(即对抗性示例、数据中毒、海绵示例)具有健壮性、能够解释其决策或保证被遗忘权。然而，这些属性相互作用的方式仍然很难确定。这对开发可信赖的机器学习模型来说是一个巨大的实际挑战。我的研究的长期愿景是创建一个框架来研究机器学习的可信性属性之间的紧张和收敛，并使用从这项研究中获得的见解来开发可以同时展示几个可信性属性的机器学习模型。更准确地说，我的研究项目将探索三个可以观察到这种紧张和趋同的重要背景。第一个背景涉及特别会议后的可解释性。在这种背景下，我建议研究后自组织解释技术的安全和隐私漏洞，并开发不受其影响的可信解释方法。第二个背景涉及同时学习公平和隐私保护模型的算法设计。在这种背景下，我提出了在数据最小化和差异隐私约束下的公平决策的新的训练技术。最后，第三个背景与机器遗忘有关。在这种背景下，我的研究将集中于研究机器遗忘中数据隐私和被遗忘权之间的权衡，以及设计健壮的数据删除算法。该研究项目的成果对学术界、行业从业者和社会都很重要，因为它们将有助于人工智能系统在伦理上保持一致的发展。它们对值得信赖的机器学习领域，更广泛地说，对于科学和工程都具有重要意义。这一研究计划的核心好处有两个：(1)洞察机器学习的可信性属性之间的紧张和趋同；(2)开发可扩展和可信任的机器学习模型的框架。此外，通过这一研究计划培训的高素质人员将帮助加拿大公司发展值得信赖的机器学习方面的专业知识。

As machine learning models become ubiquitous in our daily lives, concerns about associated ethical issues are becoming increasingly important. As a result, we are witnessing the emergence of many public and private initiatives, calling for an ethically aligned development of artificial intelligence. Machine learning models are expected to exhibit several trustworthiness properties such as being fair, robust against privacy attacks (i.e., membership inference, property inference, model extraction) and security vulnerabilities (i.e., adversarial examples, data poisoning, sponge examples), able to explain their decisions, or to guarantee the right to be forgotten. However, the way in which these properties interact with each other remains very poorly characterized. This represents a substantial practical challenge to the development of trustworthy machine learning models. The long-term vision of my research is to create a framework to study the tensions and convergences between the trustworthiness properties of machine learning and use the insights gained from this study to develop machine learning models that can simultaneously exhibit several trustworthiness properties. More precisely, my research program will explore three important contexts where such tensions and convergences can be observed. The first context concerns post-hoc explainability. In this setting, I propose studying the security and privacy vulnerabilities of post-hoc explanation techniques and developing trustworthy explanation methods that are immune to them. The second context concerns the design of algorithms to learn simultaneously fair and privacy-preserving models. In this setting, I propose new training techniques for fair decision-making under data minimization and differential privacy constraints. Finally, the third context concerns machine unlearning. In this setting, my research will focus on studying the trade-offs between data privacy and the right to be forgotten in machine unlearning and the design of robust data deletion algorithms. The outcomes of this research program are important for the academic community, industry practitioners, and society, as they will contribute to an ethically aligned development of artificial intelligence systems. They are significant for the trustworthy machine learning field and, more generally, for science and engineering. The core benefits of this research program are twofold: (1) insights into the tensions and convergences between trustworthiness properties of machine learning and (2) a framework to develop scalable and trustworthy machine learning models. Furthermore, the highly qualified personnel trained through this research program will help Canadian companies develop expertise in trustworthy machine learning.