Linyi Li 李林翼 Second Year PhD Student, CS@UIUC


I am Linyi Li, a second-year PhD Student at Department of Computer Science, University of Illinois at Urbana-Champaign. I am co-advised by Bo Li and Tao Xie.

Currently, I am working on secure machine learning, especially training certifiable neural networks and verifying neural networks. Besides, I have general interests on theory, software engineering, and programming languages.

I got bachelor degree from Department of Computer Science and Technology, Tsinghua University, where I did research on Web API Automated Testing, advised by Xiaoying Bai.

[Curriculum Vitae]



  1. Linyi Li*, Zexuan Zhong*, Bo Li, Tao Xie
    Robustra: Training Provable Robust Neural Networks over Reference Adversarial Space
    International Joint Conference on Artificial Intelligence(IJCAI) 2019
    [Paper]   [BibTex]   [Code]
  2. Klas Leino, Shayak Sen, Anupam Datta, Matt Fredrikson, Linyi Li
    Influence-Directed Explanations for Deep Convolutional Networks
    International Test Conference(ITC) 2018; Arxiv Preprint 1802.03788
    [Paper]   [BibTex]
  3. Junyi Wang, Xiaoying Bai, Linyi Li, Haoran Ma, Zhicheng Ji
    A Model-Based Framework For Cloud API Testing
    Computer Software and Applications Conference (COMSPAC), 2017 IEEE 41st Annual
    [Paper]   [BibTex]
  4. Junyi Wang, Xiaoying Bai, Haoran Ma, Linyi Li, Zhicheng Ji
    Cloud API Testing
    Verification and Validation Workshops (ICSTW), 2017 IEEE International Conference on Software Testing
    [Paper]   [BibTex]



Robustra: Training Provable Robust Neural Networks over Reference Adversarial Space

Robustra is an approach for training provable robust neural networks. The key innovation is that, instead of training over the whole perturbation space, we mutually train a pair of models on the adversarial space of the other model. By narrowing the space we improve the optimization effect and yield more robust neural networks. Particularly, on MNIST with epsilon = 0.1 l-infty ball, we reduce the provable error bound to 2.09%.

Paper Code

Lapis: Scenario-Based Automatic Web API Testing

Lapis is an automatic scenario-based Web API tester. The tool reads OpenAPI specification script and scenario definition, then generates and executes test cases automatically. Several evaluation experiments reveal its high efficiency and strength in Web API testing. PyLapis, the latest tool written in Python, using specification language extended from OpenAPI 3.0, is about to release.

Neural Network Explanation

Application of Integrated Gradients on Diabetic Retinopathy Detection Network

The project applies integrated gradients, an influence analysis method, to a diabetic retinopathy detection convolutional neural network. The tool and framework supports multiple explaining configurations such as direct attributing and middle-layer filtered attributing. The attribution results can be used directly for lesion detection. The visualization result of each neuron's influence enabled further analysis of the neural network.



A distributed web API testing system. Distributed cluster nodes send test request individually under center control. The distribution property makes it possible to generate heavy test load.

Ray Tracing

Ray Tracing Render Engine

  • A totally independent cross-platform graphics render engine.
  • Supported algorithms: Phong reflection model, ray tracing and photon mapping.
  • Supported light source: point light, and area light.
  • Supported model: DSL-specified model, and ".obj" format model.
  • Supported material: solid, and transparency with refraction.


Last Updated: Oct 11, 2019