Edge computing is a paradigm that brings computing closer to where the data is generated, such as cellular networks, internet service providers, and enterprise networks. Typical edge nodes are static and are located in fixed points over a wired network. This project creates a new notion of edge computing, Nautilus, to support distributed machine learning over mobile nodes, such as vehicles. The project aims to create new techniques that can enable a new class of applications that leverage sensors mounted on vehicles. The applications range from smart transportation, to urban planning, and to broadly support smarter cities and communities.

This project is structured to create a three-tiered architecture consisting of the nomadic edge (in vehicles), the static edge (co-located with Radio Access Networks), and the cloud. To support diverse emerging applications across domains (e.g., smart transportation, urban planning, and more broadly across different aspects of smarter communities) which can issue queries spanning large spatio-temporal regions. The research agenda includes five complementary tasks: (i) Design of distributed and dynamic computer vision from the vehicular context, which involves collaborative and federated machine learning approaches to event and object detection under diverse conditions; (ii) Design of collaborative and distributed training and inference to address high level queries which will utilize various techniques of redundant and coded computing and communication to support efficiency and scalability; (iii) Estimation and prediction of the wireless network context to infer opportunities for communication between collaborating nomadic edge nodes, as well as between the nomadic and static edges; (iv) Design for privacy and security of both data and models; and (v) Systems integration and field trials to allow for technique refinement, evaluation, and reproducibility.

• Suman Banerjee
• Mohit Gupta
• Kangwook Lee
• Kassem Fawaz

• Shenghong Dai
• Tuan Dinh
• Bhavya Goyal

* Network Side Digital Contact Tracing on a Large University Campus Matthew Malloy, Lance Hartung, Steven Wangen, Suman Banerjee ACM MobiCom, 2022

*Breaking Fair Binary Classification with Optimal Flipping Attacks C. Jo, J. Sohn, and K. Lee ISIT 2022

*Debiasing Pre-Trained Language Models via Efficient Fine-tuning M. Gira, R. Zhang, and K. Lee ACL Workshop on Language Technology for Equality, Diversity, Inclusion 2022

*Federated Unsupervised Clustering with Generative Models J. Chung, K. Lee, and K. Ramchandran AAAI Workshop on Federated Learning 2022

*Improving Fairness via Federated Learning Y. Zeng, H. Chen, and K. Lee AAAI Workshop on Federated Learning 2022

Geometric Calibration of Single-Pixel Distance Sensors Carter Sifferman, Mohit Gupta, Michael Gleicher Robotics and Automation Letters (RAL); IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) 2022

* Single-Photon Structured Light Varun Sundar, A Sankaranarayanan, Mohit Gupta Proc. IEEE CVPR 2022

* Compressive Single-Photon 3D Cameras Felipe Gutierrez-Barragan, Atul Ingle, T Seets, Mohit Gupta, Andreas Velten Proc. IEEE CVPR 2022

* Single-Photon Camera Guided Extreme Dynamic Range Imaging Yuhao Liu, Felipe Gutierrez-Barragan, Atul Ingle, Mohit Gupta, Andreas Velten IEEE Winter Conference on Applications of Computer Vision (WACV 2022)

Sample Selection for Fair and Robust Training Y. Roh, K. Lee, S. Whang, and C. Suh, NeurIPS 2021.

Gradient Inversion with Generative Image Prior J. Kim, J. Jeon, K. Lee, S. Oh, and J. Ok, NeurIPS 2021.

A General Framework For Detecting Anomalous Inputs to DNN Classifiers J Raghuram, V Chandrasekaran, S Jha, S Banerjee, International Conference on Machine Learning, 2021.

All Roads Lead to Rome: An MPTCP-Aware Layer-4 Load Balancer Y Zeng, M Buddhikot, S Banerjee IFIP Networking Conference (IFIP Networking), 2021.

* S. Ahmed, I. Shumailov, N. Papernot, and K. Fawaz, Towards More Robust Keyword Spotting for Voice Assistants, in 31st USENIX Security Symposium (USENIX Security 2022), Accepted.

* Photon Starved Scene Inference using Single-Photon Cameras. Bhavya Goyal and Mohit Gupta. Proc. IEEE International Conference on Computer Vision (ICCV 2021).

* Coded-InvNet for Resilient Prediction Serving Systems T. Dinh and K. Lee ICML 2021 (long oral)

* Powercut and obfuscator: An exploration of the design space for privacy-preserving interventions for smart speakers. V. Chandrasekaran, S. Banerjee, B. Mutlu, and K. Fawaz. In Seventeenth Symposium on Usable Privacy and Security (SOUPS 2021). USENIX Association, Aug. 2021. [Online]. Available: https://www.usenix.org/conference/soups2021/presentation/chandrasekaran

* (June 2021) Co-PI Lee gave an invited talk @ AI institute of POSTECH on FairBatch and Coded-InvNet

* (June 2021) Co-PI Lee gave an invited talk at the Shannon meets Turing Colloquium @ Seoul National University on FairBatch and Coded-InvNet

* Anant Gupta, Atul Ingle, and Mohit Gupta. “Asynchronous single-photon 3D imaging.” Proc. of the IEEE/CVF International Conference on Computer Vision. 2019.

* Gupta, A., Ingle, A., Velten, A., & Gupta, M. (2019). Photon-flooded single-photon 3D cameras. Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2019.

* Ingle, A., Velten, A., & Gupta, M. (2019). High flux passive imaging with single-photon sensors. Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2019.

* FairBatch: Batch Selection for Model Fairness. Y. Roh, K. Lee, S. Whang, and C. Suh. ICLR 2021 https://openreview.net/pdf?id=YNnpaAKeCfx

* Accordion: Adaptive Gradient Communication via Critical Learning Regime Identification. S. Agarwal, H. Wang, K. Lee, S. Venkataraman, and D. Papailiopoulos. MLSys 2021 https://proceedings.mlsys.org/paper/2021/hash/1d7f7abc18fcb43975065399b0d1e48e-Abstract.html

N/A

* (April 2021) Co-PI Lee's paper “Accordion” is presented at MLSys 2021

* (May 2021) Co-PI Lee's paper “FairBatch” will be presented at ICLR 2021

* Shenghong Dai is implementing the mobile learning agents in the vehicle simulator CARLA. Her simulator will play a key role in testing the realistic federated learning with mobile agents.

* Tuan Dinh is working on finalizing the paper on coded computation + invertible neural networks. In particular, he is working on the use of his coded computation algorithm for normalized flows for generative learning.

N/A

* Co-PI Lee will give one invited talk on fairness issues in machine learning: i) (April 16th) FairBatch @ the special interest group (SIG) at the intersection of ethics and algorithms research composed of researchers from UC Santa Cruz, The University of Wisconsin at Madison, and The University of Washington, and the University of Chicago

Nothing noted.

* Attack of the Tails: Yes, You Really Can Backdoor Federated Learning. H. Wang, K. Sreenivasan, S. Rajput, H. Vishwakarma, S. Agarwal, J. Sohn, K. Lee, and D. Papailiopoulos. NeurIPS 2020 https://papers.nips.cc/paper/2020/file/b8ffa41d4e492f0fad2f13e29e1762eb-Paper.pdf

* Face-Off: Adversarial Face Obfuscation. Varun Chandrasekaran, Chuhan Gao, Brian Tang, Kassem Fawaz, Somesh Jha, Suman Banerjee. PoPETS 2021. https://arxiv.org/abs/2003.08861

* PriSEC: A Privacy Settings Enforcement Controller. R. Khandelwal, T. Linden, H. Harkous, and K. Fawaz. Accepted in 30th USENIX Security Symposium (USENIX Security 2021). https://www.usenix.org/system/files/sec21summer_khandelwal.pdf

* Kaleido: Real-Time Privacy Control for Eye-Tracking Interactive Systems. J. Li, A. R. Chowdhury, Y. Kim, and K. Fawaz. Accepted in 30th USENIX Security Symposium (USENIX Security 2021). https://www.usenix.org/system/files/sec21summer_li-jingjie.pdf

* Poster presented at NeurIPS 2020

* Employing redundancy in the feature space and physical space to improve the robustness of keyword spotting systems. We have begun developing an approach to distribute the sensing of wake words in an environment. The distributed sensing leverages multiple devices and microphones as well as multiple model architectures to improve the robustness of keyword spotting against adversarial examples.

* Coded-InvNet for Resilient Prediction Serving Systems Our experimental results show that Coded-InvNet can outperform existing approaches, especially when the compute resource overhead is as low as 10%. For instance, without knowing which of the ten workers is going to fail, our algorithm can design a backup task so that it can correctly recover the missing prediction result with a probability of 85.9%, significantly outperforming the previous SOTA 53.4%

* Federated Learning with Mobile Learning Agents

* Distributed 3D Cameras with low-cost proximity sensors. We have begun developing a theoretical framework that allows using multiple ultra-low-cost single-pixel proximity sensors as a single distributed 3D camera. The key technical challenge is determining the relative poses and locations of individual sensors. Our initial results show that it is possible to perform such calibration for a certain families of sensor configurations. Next steps involve considering a more general set of spatial configurations.

* Co-PI Lee's work received the Joint Communications Society/Information Theory Society Paper Award. The idea developed in this paper is one of the key technical components of our project. K. Lee, M. Lam, R. Pedarsani, D. Papailiopoulos, and K. Ramchandran, “Speeding up Distributed Machine Learning Using Codes”, IEEE Transactions on Information Theory, vol. 64, no. 3, pp. 1514-1529, March 2018

* Co-PI Lee gave three invited talks on fairness issues in machine learning: i) BLISS seminar @ UC Berkeley ii) SILO seminar @ UW Madison iii) ML Ideas @ Microsoft Research New England

Nothing noted.