CPS: Medium: Secure Constrained Machine Learning for Critical Infrastructure CPS

机器学习在计算机视觉、语音分析和自然语言处理等现代商业应用领域取得了许多成功。然而，它在关键基础设施网络物理系统（CI-CPS）中的广泛应用，如能源、水、交通、石油和天然气系统，远不理想。这主要是由于对现有机器学习技术的可靠性和学习模型缺乏可解释性的担忧。此外，CI-CPS经常直接从商业应用程序中借用技术，而这些应用程序没有考虑这些系统中固有的物理和拓扑约束。机器学习的安全性最近得到了广泛的研究，揭示了机器学习模型的漏洞以及通过污染模型输入来偏离学习结果的有效性。这在CI-CPS中尤其具有破坏性，因为学习用于安全关键操作，这种偏差可能对人员和物理资产造成不可逆转的伤害。因此，建立独特的CI-CPS约束模型的安全机器学习是一个非常需要的研究领域，也是本项目的重点。该提案涉及安全、机器学习和CI-CPS三个领域，旨在提高现代社会重要基础设施的安全性和弹性。我们使用两个CI-CPS，电力系统和交通系统作为目标应用领域，以说明所提出方法的一般适用性。提出的工作由四个研究任务进行。首先，该项目将设计一个合适的威胁模型，在该模型下，对抗性机器学习攻击（ConAML）在CI-CPS约束下开发。其次，该项目将通过在训练和推理中引入随机输入填充来提出一种缓解ConAML攻击的方法。第三，该项目将提出一个新的“数据表示-模型-任务”关联框架，通过设计一个变异的dirichlet网络，在表示空间而不是原始数据空间中将输入数据与机器学习模型桥接起来，从底层实现安全的约束机器学习。最后，项目团队将提出的安全约束机器学习应用于电力负荷预测和流量预测，在测试平台中实现这些应用程序，并评估其在ConAML攻击下的安全性和性能。提出的研究旨在提高CI-CPS的安全性、可靠性和弹性。它为CI-CPS的安全机器学习知识库做出了贡献，并适用于所有安全关键型大型互联CPS。拟议工作的多学科性质有助于未来科学家和工程师的跨学科教育和培训。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Machine learning has found many successes in modern commercial application domains like computer vision, speech analysis, and natural language processing. However, its broader use in critical infrastructure cyber-physical systems (CI-CPS), such as, energy, water, transportation, and oil and natural gas systems, has been far less than ideal. This is mainly due to concerns with the reliability of existing machine learning techniques and the lack of explainability of the learned models. Moreover, CI-CPS often borrow techniques directly from commercial applications that fail to consider physical and topological constraints inherent in these systems. Security of machine learning has been extensively studied recently, revealing vulnerabilities of machine learning models and the effectiveness in deviating learning outcomes by polluting the model input. This is especially devastating in CI-CPS where learning is used for safety-critical operations and such deviation can cause irreversible harm to people and physical assets. Secure machine learning that models unique CI-CPS constraints is thus a much needed research area and is the focus of this project. This proposal intersects three fields - security, machine learning, and CI-CPS - to enhance the safety and resiliency of essential infrastructures in modern society. We use two CI-CPS, power systems and transportation systems, as target application domains to illustrate the general applicability of the proposed approach. The proposed work is carried out by four research tasks. First, the project will devise a suitable threat model under which adversarial machine learning attacks, ConAML, are developed subject to CI-CPS constraints. Second, the project will propose a mitigation method for ConAML attacks by introducing random input padding in both training and inference. Third, the project will propose a new “data-representation-model-task” association framework that realizes secure constrained machine learning from ground up, by designing a variation Dirichlet-network that bridges the input data with machine learning models in the representation space instead of the raw data space. Lastly, the project team will apply the proposed secure constrained machine learning to electric load forecasting and traffic forecasting, implement these applications in testbeds, and evaluate their security and performance under ConAML attacks. The proposed research seeks to improve the security, reliability and resiliency of CI-CPS. It contributes to the knowledge base of secure machine learning for CI-CPS, and applies to all safety-critical large interconnected CPS. The multi-disciplinary nature of the proposed work lends itself to cross-disciplinary education and training of future scientists and engineers.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.