Collaborative Research: NSF-BSF: SaTC: CORE: Small: Detecting malware with machine learning models efficiently and reliably

检测恶意软件（又名恶意软件）是机器学习（ML）越来越多地用于执行的常见安全挑战。虽然存在许多类型的恶意软件检测器，但本项目侧重于改进那些无需实际运行即可识别恶意软件的检测器。在这些类型的检测器中，机器学习模型在许多恶意软件和良性软件的示例上进行训练，以学习如何根据它们包含的计算机指令中的模式对新示例进行分类。使用这种机器学习模型的一个特定威胁是，攻击者可能会利用他们对模型工作原理的理解来开发恶意软件变体，这些变体对模型来说尤其难以检测。该项目旨在开发用于恶意软件检测的机器学习模型，尽管攻击者试图欺骗它们，但这些模型仍然可靠，同时开发出更节省时间和空间的模型。该项目的技术创新在于对技术的基础研究，以比迄今为止可能的更高的效率，定量地提高恶意软件检测器对被愚弄的抵抗力。这些技术包括将多个分类器组合成集成的新方法；新颖的模型架构；利用将分类结果归因于输入的特定属性的方法；以及对恶意软件领域的机器学习训练方法的改进。该项目将开发这些技术应用的基本原则，以改进机器学习模型，并通过使用恶意软件和良性可执行文件的数据集，实证地展示它们产生的改进。如果这样做是负责任的（即，不会冒险促进对操作系统的新攻击），这个项目的进展将开放源代码，供其他人在后续研究中利用；在其他情况下，实现将仅根据请求分发给可验证的反恶意软件工具的研究人员或开发人员。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。

Detecting malicious software (a.k.a. malware) is a common security challenge that machine learning (ML) is increasingly used to perform. While many types of malware detectors exist, this project focuses on improving those that identify malware without actually running it. In these types of detectors, an ML model is trained on many examples of malware and of benign software to learn how to classify new examples based on patterns in the computer instructions they contain. A specific threat to the use of such ML models is that attackers might use their understanding of a model's workings to develop malware variants that are particularly hard for the model to detect. This project seeks to develop ML models for malware detection that are reliable despite attackers' attempts to fool them, as well developing models that are more time- and space-efficient.The technical innovations of this project lie in basic research on techniques to quantifiably improve the resistance of malware detectors to being fooled, with greater efficiency than has been possible to date. These techniques include new approaches to combining multiple classifiers into ensembles; novel model architectures; leveraging methods that attribute classification outcomes to specific properties of the input; and improvements to ML training methods for the malware domain. This project will develop principles underlying the application of these techniques to improve ML models and empirically demonstrate the improvements they yield, using datasets of malware and benign executables. Where doing so is responsible (i.e., does not risk facilitating new attacks on operational systems), the advances of this project will be open-sourced for others to leverage in follow-on research; in other cases, the implementations will be distributed only by request to parties who are verifiable researchers or developers of anti-malware tools.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.