Model Compression for Object Tracking           (DARPA IP2 Program)

Supervisor: Sijia Liu (MSU)
Sept. 2021 - Present

  • Propose a hardware-friendly pruning scheme for the task of object tracking

  • Adopt knowledge distillation to acquire lightweight and high-accuracy model

  • Achieve 90% model sparsity without performance loss for ResNet-50 under BDD100K dataset

Robustification of Black-Box ML Models by Zeroth-Order Optimization                      

Supervisor: Sijia Liu (MSU)      Collaborator: Jinfeng Yi (JD AI), Mingyi Hong (UMN), Shiyu Chang (UCSB)
Jan. 2021 - Oct. 2021

  • Formulate black-box defense problem through the lens of zeroth-order (ZO) optimization

  • Propose scalable ZO optimization method to tackle defense challenge in high dimension

  • Achieve state-of-the-art certified robustness on CIFAR-10 and STL-10

  • Extend black-box defense from image classification to image reconstruction

  • Publication: Zhang, Y., Yao, Y., Jia. J., Yi, J., Hong, M., Chang, S., Liu, S. How to Robustify Black-Box ML Models? A Zeroth-Order Optimization Perspective, International Conference on Learning Representation (ICLR’22 - Spotlight)

RED: Reverse Engineering of Deceptions           (DARPA RED Program)

Supervisor: Sijia Liu (MSU)      Collaborator: Xiaoming Liu (MSU), Xue Lin (NEU)
Mar. 2021 - Oct. 2021

  • Design Reverse Engineering of Deceptions (RED) pipeline to recover adversarial perturbations

  • Integrating RED with data augmentation techniques to overcome unforeseen attacks

  • Identify RED principles: pixel-level reconstruction, prediction-level alignment, and attributionlevel saliency recovery

  • Publication: Gong, Y., Yao, Y., Li, Y., Zhang, Y., Liu, X., Lin, X., Liu, S. Reverse Engineering of Imperceptible Adversarial Image Perturbations, International Conference on Learning Representation (ICLR’22)

Video Synthesis via Transform-Based Tensor Neural Network

Supervisor: Anwar Walid (Columbia University)
Aug. 2019 - May 2020

  • Propose an iterative tensor ISTA algorithm for video processing

  • Design a Transform-Based Tensor-Net for video frame synthesis task

  • Achieve state-of-the-art PSNR on KTH and UCF-101

  • Publication: Zhang, Y., Liu, X. Y., Wu, B., & Walid, A. Video Synthesis via Transform-Based Tensor Neural Network, ACM International Conference on Multimedia (ACM MM’20)

Tensor FISTA-Net for Real-Time Snapshot Compressive Imaging

Supervisor: Linghe Kong (SJTU)
April. 2019 - Oct. 2019

  • Propose a novel Tensor FISTA-Net for SCI reconstruction

  • Utilize tensor form to reduce time and memory consumption significantly

  • Achieve state-of-the-art reconstruction accuracy and speed on both synthetic and real datasets

  • Small model size (12MB) makes it practical for real-time IoT applications

  • Publication: Han, X., Wu, B., Shou, Z., Liu, X. Y., Zhang, Y., Kong, L. Tensor FISTA-Net for real-time snapshot compressive imaging, AAAI Conference on Artificial Intelligence (AAAI’20)