SHF: Small: Software and Hardware Support for Robust Deep Learning

近年来，深度神经网络取得了前所未有的进步，使其成为任何现代计算系统中无处不在的组件。深度神经网络及其硬件将在未来的各种系统中司空见惯，从简单的任务到安全和安保关键系统。因此，必须有专门的支持来提高深度神经网络模型的稳健性，以便用户和开发人员能够信任这些模型提供的预测。为此，本项目提出了一种新的算法来实现稳健的深层神经网络，旨在确保深层神经网络的稳健运行。该项目名为GuardiaNN，由两部分组成--动态监测方案和测试方案。动态监控的目标是确定深度神经网络模型在生产环境中的推理过程中是否会出现错误行为。该项目考虑了两种新的监测算法--神经元模式监测和基于模型的监测。基于神经元模式的方案的工作原理是，首先提取简单的神经元激活模式，然后在运行时检查这些模式是否满足。基于模型的方案的工作原理是，首先确定要监控的最小数据点集合，然后使用高级模型来检测这些数据点是否显示任何异常模式。测试方案的目标是确定模型在角落情况下可能的行为。测试允许用户在不了解模型的内部细节的情况下探索模型的行为。监控和测试方法将提供对深度神经网络操作的信任。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

Deep neural networks have made unprecedented strides in recent years, making them a ubiquitous component of any modern computing system. Deep neural networks and their hardware will be commonplace in future in various systems ranging from simple tasks to safety and security-critical systems. Therefore, it is imperative to have specialized support to improve the robustness of deep neural network models so that users and developers can trust the predictions provided by those models. Towards that end, this project presents novel algorithms for robust deep neural networks.The project aims at ensuring robust operations of deep neural networks. The project, called GuardiaNN, has two components - dynamic monitoring schemes and testing schemes. The goal of dynamic monitoring is to determine whether a deep neural network model can misbehave during inference at a production setting. Two novel monitoring algorithms are considered in this project - neuron pattern and model-based monitoring. Neuron pattern-based schemes work by first extracting simple neuron activation patterns and then, later at runtime, checking whether those patterns are satisfied or not. Model-based schemes work by first determining a minimal set of data points to monitor and then, using an advanced model to detect if those data points show any anomalous pattern or not. The goal of testing schemes is to determine how the model might behave in corner case scenarios. Testing allows users to explore the behavior of the model without understanding the internal details of the model. Together the monitoring and testing approaches will provide trust in deep neural network operations.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.