March 2022

Three papers in three premier conferences, respectively in security (IEEE Security and Privacy), machine learning/computer vision (CVPR), and systems (Eurosys).

Congratulations to Mustafa for leading the charge on S&P and Ran for CVPR and Eurosys. Also congratulations to Daniel for contributing to S&P, Jay and Preeti to CVPR, and Jay and Pengcheng to Eurosys.

The S&P paper is the result of collaboration with Economics colleague, Tim (Cason) and ECE colleague, Shreyas (Sundaram) and involves 2 other organizations — Parinaz (Naghizadeh) who is at Ohio State and Issa (Khalil) who is at Qatar Computing Research Association (QCRI).

The CVPR and Eurosys papers are from collaboration with CV expert, Yin (Li) from University of Wisconsin at Madison, and Purdue colleague, Somali (Chaterji). Fangzhou from Wisconsin is a key contributor as well.

Mustafa Abdallah, Daniel Woods; Parinaz Naghizadeh (Ohio State University); Issa Khalil (Qatar Computing Research Institute (QCRI), HBKU); Timothy Cason, Shreyas Sundaram, Saurabh Bagchi, “TASHAROK: Using Mechanism Design for Enhancing Security Resource Allocation in Interdependent Systems,” Accepted to the 43rd IEEE Symposium on Security and Privacy (S&P 2022).

We consider interdependent systems managed by multiple defenders that are under the threat of stepping-stone attacks. We model such systems via game-theoretic models and incorporate the effect of behavioral probability weighting that is used to model biases in human decision-making, as descended from the field of behavioral economics. We then incorporate into our framework called TASHAROK, two types of tax-based mechanisms for such interdependent security games where the central regulator incentivizes defenders to invest well in securing their assets so as to achieve the socially optimal outcome. We first show that due to the nature of our interdependent security game, no reliable tax-based mechanism can incentivize the socially optimal investment profile while maintaining a weakly balanced budget (in such a budget, the central regulator does not have to make net payouts). We then show the effect of behavioral probability weighting bias on the amount of taxes paid by defenders and prove that higher biases make defenders pay more taxes under the two mechanisms. We then explore voluntary participation in tax-based mechanisms. To evaluate our mechanisms, we use four representative real-world interdependent systems where we compare the game-theoretic optimal investments to the socially optimal investments under the two mechanisms. We show that the mechanisms yield a higher decrease in the social cost for behavioral decision-makers compared to rational decision-makers.

Ran Xu, Jayoung Lee, Pengcheng Wang, Saurabh Bagchi; Yin Li (University of Wisconsin – Madison); Somali Chaterji, “LiteReconfig: Cost and Content Aware Reconfiguration of Video Object Detection Systems for Mobile GPUs,” Accepted to appear at the 17th European Conference on Computer Systems (EuroSys 2022).

An adaptive video object detection system selects different execution paths at runtime, based on a user specified latency requirement, video content characteristics, and available resources on a platform, so as to maximize its accuracy under the target latency service level agreement (SLA). Such a system is well suited for mobile devices with limited computing resources, often times running multiple contending applications. In spite of several recent efforts, we show that existing solutions suffer from two major drawbacks when facing a tight latency requirement (e.g., 30 fps). First, it can be very expensive to collect some feature values for a scheduler to decide on the best execution branch to run. Second, the system suffers from the switching overhead of transitioning from one branch to another, which is variable depending on the transition pair. This paper addresses these challenges and presents LiteReconfig — an efficient and adaptive video object detection framework for mobiles. Underlying LiteReconfig is a cost-benefit analyzer for the scheduler that decides which features to use and then which execution branch to run at inference time. LiteReconfig is further equipped with a content-aware accuracy prediction model to select an execution branch tailored for frames in a streaming video. With a large-scale real-world video dataset and multiple current generation embedded devices, we demonstrate that LiteReconfig achieves significantly better accuracy under a set of varying latency requirements when compared to existing adaptive object detection systems, while running at speeds up to 50 fps on an NVIDIA AGX Xavier board.

Ran Xu; Fangzhou Mu (University of Wisconsin-Madison); Jayoung Lee,
Preeti Mukherjee, Somali Chaterji, Saurabh Bagchi; Yin Li (University of Wisconsin-Madison), “SMARTADAPT: Multi-branch Object Detection Framework for Videos on Mobiles,” Accepted to appear at the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR 2022).

Several recent works seek to create lightweight deep networks for object detection on mobiles and, in particular, for video object detection. Many existing detectors intrinsically support adaptive inference, and offer a multi-branch object detection framework (MBODF). Here, an MBODF is referred to as a solution that has many execution branches and one can dynamically choose from among them at inference time to satisfy varying latency requirements (e.g., by varying resolution of the input frame). In this paper, we ask, and answer, the wide-ranging question across all MBODFs: How to expose the right set of execution branches and then how to schedule the optimal one at inference time? In addition, we uncover the importance of making a content-aware decision on the branch to run, as the optimal one is conditioned on the video content. Finally, we explore a content-aware scheduler, an Oracle one, and then a practical one, leveraging various lightweight feature extractors. Our evaluation shows that layered on Faster R-CNN-based MBODF, compared to 7 baselines, our method SMARTADAPT achieves a higher Pareto optimal curve in the accuracy-vs-latency space for ILSVRC VID dataset.