Zhilu Wang's research while affiliated with Northwestern University and other places

Publications (24)

Preprint
With the increment of interest in leveraging machine learning technology in safety-critical systems, the robustness of neural networks under external disturbance receives more and more concerns. Global robustness is a robustness property defined on the entire input domain. And a certified globally robust network can ensure its robustness on any pos...
Preprint
The robustness of deep neural networks has received significant interest recently, especially when being deployed in safety-critical systems, as it is important to analyze how sensitive the model output is under input perturbations. While most previous works focused on the local robustness property around an input sample, the studies of the global...
Article
During the operation of many real-time safety-critical systems, there are often strong needs for adapting to a dynamic environment or evolving mission objectives, e.g., increasing sampling and control frequencies of some functions to improve their performance under certain situations. However, a system's ability to adapt is often limited by tight r...
Preprint
In the current control design of safety-critical autonomous systems, formal verification techniques are typically applied after the controller is designed to evaluate whether the required properties (e.g., safety) are satisfied. However, due to the increasing system complexity and the fundamental hardness of designing a controller with formal guara...
Preprint
Neural networks are being increasingly applied to control and decision-making for learning-enabled cyber-physical systems (LE-CPSs). They have shown promising performance without requiring the development of complex physical models; however, their adoption is significantly hindered by the concerns on their safety, robustness, and efficiency. In thi...
Article
With growing system complexity and closer cyber-physical interaction, there are increasingly stronger dependencies between different function and architecture layers in automotive systems. This paper first introduces several cross-layer approaches we developed in the past for holistically addressing multiple system layers in the design of individua...
Preprint
Many safety-critical real-time systems operate under harsh environment and are subject to soft errors caused by transient or intermittent faults. It is critical and yet often very challenging to apply fault tolerance techniques in these systems, due to their resource limitations and stringent constraints on timing and functionality. In this work, w...
Preprint
In many intelligent systems, a network of agents collaboratively perceives the environment for better and more efficient situation awareness. As these agents often have limited resources, it could be greatly beneficial to identify the content overlapping among camera views from different agents and leverage it for reducing the processing, transmiss...
Preprint
Full-text available
With growing system complexity and closer cyber-physical interaction, there are increasingly stronger dependencies between different function and architecture layers in automotive systems. This paper first introduces several cross-layer approaches we developed in the past for holistically addressing multiple system layers in the design of individua...
Preprint
Full-text available
Control schemes for autonomous systems are often designed in a way that anticipates the worst case in any situation. At runtime, however, there could exist opportunities to leverage the characteristics of specific environment and operation context for more efficient control. In this work, we develop an online intermittent-control framework that com...
Chapter
In many cyber-physical systems (CPS), software has become critical and drives future innovations. CPS software development, however, faces significant challenges from increasing functional and architectural complexity, dynamic and uncertain physical environment, and diverse design objectives and stringent system requirements. In this chapter, we in...
Conference Paper
The design of automotive electronic systems needs to address a variety of important objectives, including safety, performance, fault tolerance, reliability, security, extensibility, etc. To obtain a feasible design, timing constraints must be satisfied and latencies of certain functional paths should not exceed their deadlines. From functionality p...

Citations

... Global robustness can perfectly avoid this issue as it can be certified offline and applied to the entire input domain. [1] is the first efficient global robustness certification algorithm that can handle neural networks with thousands of neurons. This work is an extension of the neural network global robustness certification algorithm [1]. ...
... It is critical yet challenging to formally ensure their safety, especially for the control and decision making modules. Thus, while there has been increasing interest in applying machine learning techniques (especially neural network based ones such as deep reinforcement learning [3]) to control and general decision making, their adoption in safety-critical systems is hindered by the challenges in formally ensuring system properties [4,5]. ...
... It measures the worst-case output variation when there is a small input perturbation for any possible input sample in the input domain X. Moreover, we focus on the input perturbation that is bounded in the form of L ∞ norm, and have the same global robustness definition as [2], [13]. ...
... Moreover, compared with traditional model-based approaches, they can save the time and effort of explicitly modeling systems with complex dynamics and significant uncertainties. However, a major challenge for the neural network based planners is to ensure system safety, especially for safety-critical applications [39] and in near-accident scenarios [1,7], such as unprotected left turn and highway merging in autonomous driving. In those scenarios, with only minor changes in environment states, dramatically different behaviors may need to be performed to avoid accidents, which is difficult for both humans and autonomous systems to handle. ...
... In systems like self-driving cars and smart robotics, various DNNs are used collectively to achieve intelligent features such as object detection, scene recognition, and natural language processing. As DNNs are increasingly involved in autonomous decision-making processes, a faulty or late output may lead to catastrophic behavior, which is unacceptable in safety-critical domain [19,40,45]. Therefore, it is an emerging challenge to ensure highly dependable real-time execution of DNN tasks with limited computing resources. ...
... This model is equivalent to the case where deadline misses represent discarded computations, but its results can not be generalized for the other common case of late completions. In other works [23], [28], [29], the authors have studied how to enforce safety guarantees of weakly-hard real-time controllers, with focus also on stability properties. However, such works consider only the case where a deadline miss corresponds to a discarded computation and, in [28], [29], with the additional hypothesis of a known periodic pattern of deadline hits and misses. ...
... With bounds on ∆d and other variables, the vehicle control safety can be verified based on the computation of control invariant set, similarly as in [20]. Through such invariant set based verification, we find that as long as the distance estimation error ∆d is within [−0.14, 0.14], the system will always be safe. ...
... A Software Defined Network (SDN) and cross-layer design allow users to respond to changes and adapt to them in real-time quickly. A software-defined network architecture creates a more manageable and dynamic network infrastructure [48,49]. Software-Defined Networking (SDN) enables a multi-layered network design with a high degree of flexibility and flexibility. ...
... That is task τ N is invoked at discrete sampling instants. Under the Logical Execution Time (LET) paradigm [46,47,48], which is widely adopted in CPS, the control signal that is computed based on the measurements at sampling instant k is applied to the plant at sampling instant k + 1. This means that there is a fixed samplingto-actuation delay which is equal to ∆T N . ...
... With six papers, the third-largest category is formed by studies that elicited and described design challenges for AI-based systems, e.g., by conducting empirical studies like surveys [228] or reporting industry experiences [75]. The scope of the described design challenges varies greatly and covers areas such as intelligent automotive systems [120], ML model management [177], or ML fairness [81]. The details of these challenges are explained in Section 7 (RQ4). ...