Trustworthy Machine Learning by Design

机器学习（ML）带来了一种新的计算范式：我们现在不再为我们想要解决的每个问题编写程序，而是编写一个学习算法，并为其提供从问题解决方案中预期的输入和输出示例。这从根本上改变了我们构建软件的方式，并使许多以前具有挑战性的领域取得了进展，如自动驾驶、医疗保健或计算机视觉。然而，机器学习的广泛采用引发了安全和社会问题，这些问题表明用于训练和预测的算法缺乏可信度。学习算法很容易被能够扰乱ML算法分析的数据的对手操纵。学习算法也可以延续和放大它们分析的数据中历史上发现的偏见。尽管研究界对这些挑战非常感兴趣，但这些挑战仍然没有得到解决，因为当逻辑是不透明的训练过程的结果时，要确保解决问题所遵循的逻辑满足所需（例如鲁棒性或公平性），而不是由人类程序员手工编码，这要困难得多。

Machine learning (ML) brought a new computing paradigm: rather than writing programs for each problem we'd like to solve, we now write a single learning algorithm and present it with examples of inputs and outputs that are expected from the solution to the problem. This has fundamentally changed the way we build software and enabled progress in many previously challenging areas such as autonomous driving, healthcare, or computer vision. Yet, the widespread adoption of ML raises security and societal concerns that point to the lack of trustworthiness of algorithms used for training and predicting. Learning algorithms can easily be manipulated by adversaries capable of perturbing the data that ML algorithms analyze. Learning algorithms can also perpetuate and magnify biases historically found in the data they analyze. These challenges remain unaddressed despite significant interest from the research community, because it is significantly more difficult to ensure that the logic followed to solve a problem satisfies desiderata (such as robustness or fairness) when this logic is the fruit of an opaque training process---rather than hand-coded by a human programmer. To achieve trustworthy ML, the proposed research is three-fold. First, we will shift away from the research community's focus on modifying existing ML architectures and algorithms to increase their trustworthiness. Instead, we will design them ab initio to train models that not only achieve high performance but also represent data in a different way: these representations of the data are constrained to satisfy properties (e.g., of security and privacy) that make it easier, by design, to verify that the resulting model is trustworthy. Second, our techniques will rely on these novel ML architectures to uncover the relationship between training data and predictions: this will make it easier to audit the observed behavior of a ML algorithm and obtain assurance that it is operating as expected, or when it is not, to limit the sensitivity of data collection processes to biases historically found in the data. Third, we will develop model governance mechanisms to manage ML models once they have been deployed to make predictions: this will notably inform policy discussions about intellectual property and disinformation. Our research will thus both help humans to trust ML, and ML algorithms to learn effectively and responsibly from humans. The facets of trustworthy ML that we consider are instrumental in ensuring a beneficial impact of ML, aligned with Canadian values. This will not only help with security and safety critical deployments of machine learning, such as the ones found in the cyber-physical infrastructure or the healthcare sector, it will also grow trust from the general society in applications of ML that have penetrated our daily lives: from the language models that help us type on our smartphones to the face recognition algorithms we use to board planes.