DEEL DEpendable & Explainable Learning

机器学习（ML）应用于航空航天工业的复杂问题，具有高标准的性能和安全性，需要使技术强大，易于理解，保证隐私，从而获得当局的认证。DEEL项目是法国的ITS Saint Exupéry、IVADO和加拿大的CRIAQ之间开展国际合作的结果，旨在通过几个行业参与者和几所加拿大大学之间的多学科合作，向使用机器学习技术解决加拿大航空航天工业的各种问题迈出第一步，这些大学有十几名研究人员和二十多名研究生，第一个主题涉及鲁棒性，包括开发ML方法，即使在设计过程中没有观察到的极端情况下也保持有效。为了完成这项工作，我们依赖于测量决策不确定性的方法，管理上下文变化，以及对攻击具有鲁棒性的方法。第二个主题涉及可解释性，并涉及开发向专家（设计师，机组人员）提供解释的方法，以使系统的决策或建议可以理解。本主题中使用的方法基于学习模型的透明性概念和给定决策的可解释性概念。第三个主题涉及设计隐私，涉及开发方法，确保用于设计ML模型的数据保持机密，并且不能从系统的结果或操作中重建。本主题中使用的方法基于数据中噪声的受控添加和加密数据的学习。这三个主题旨在提供提高ML系统结果的可信度并实现可认证性的技术。第四个主题的目标是开发ML认证技术。我们建议使用软件工程的方法和最佳实践，例如测试和形式化方法，来发现如何使它们适应ML。

The application of machine learning (ML) to the complex problems of the aerospace industry, which has high standards of performance and security, needs to make the techniques robust, comprehensible, guaranteeing privacy and thus certifiable by the authorities. The DEEL project, as a result of an international collaboration between ITS Saint Exupéry in France, IVADO and CRIAQ in Canada, aims to take a first step towards the use of machine learning techniques for various problems of the Canadian aerospace industry through a multidisciplinary collaboration between several industry players and several Canadian universities with more than a dozen researchers and more than twenty graduate students and highly qualified professionals annually.The first theme deals with robustness and consists of developing ML methods that remain effective even in extreme situations not observed during design. To do this work, we rely on methods for measuring decision uncertainty, managing context changes, and methods that are robust to attack. The second theme deals with interpretability and involves developing methods that provide explanations to experts (designer, crew) to make the decisions or advice of the system understandable. The methods used in this theme are based on the notion of transparency of the learned model and on the notion of explicability of a given decision.The third theme deals with privacy by design and involves developing methods that ensure that the data used to design the ML model remains confidential and cannot be rebuilt from the results or operation of the system. The methods used in this theme are based on the controlled addition of noise in data and learning on encrypted data.These three themes aim to provide techniques to improve trust in the results of ML systems and to enable certifiability. The objective of this fourth theme is to develop ML certifiability techniques. We propose using methods and best practices of software engineering, such as tests and formal methods, to discover how to adapt them to ML.