Zhiyu Ni

zhiyuni-at-usc.edu

I am a PhD student studying Electrical Engineering at the University of Southern California. I joined USC on 2022 Fall and I am currently advised by Professor Pierluigi Nuzzo, DesCyPhy Lab.
My primary research delves into trustworthy machine learning like security and privacy. I am expert at GNN. Recently I am also working on the safety of LLM and anomaly detection!

Education

University of Southern California

Los Angeles, CA, USA
Ph.D. in Computer Engineering
August 2022 - Present

University of Science and Technology of China

Hefei, Anhui, China
B.E. in Physics
Advisor: Cao Ping
August 2018 - July 2022

University of Michigan

Ann Arbor, MI, USA
Research Intern in Physics
Advisor: Junjie Zhu
September 2020 - January 2021

Publications

Sparse but Strong: Crafting Adversarially Robust Graph Lottery Tickets

Subhajit Dutta Chowdhury*, Zhiyu Ni*, Qingyuan Peng, Souvik Kundu, Pierluigi Nuzzo

Accepted by GLFrontiers Workshop, NeurIPS 2023

Finding Adversarially Robust Graph Lottery Tickets

Subhajit Dutta Chowdhury*, Zhiyu Ni*, Qingyuan Peng, Souvik Kundu, Pierluigi Nuzzo

Under review of ICLR 2024

Projects

Exploring Robustness of Graph Lottery Tickets Against Adversarial Attacks

  • Systematically analyzed the robustness of graph lottery tickets (GLT) against adversarial attacks.
  • Proposed a new loss function for training robust yet sparse GLTs against poisoning and evasion attacks.
March 2023 - September 2023

LLM for Anomaly Detection

  • Exploring the capabilities of LLMs (ChatGpt, Llama, etc.) in terms of anomaly detection and encoding graph information into natural language.
  • Designing in-context learning flow to enable LLMs to identify fraud and numerically evaluate performance in various datasets (YelpChi, AmazonChi).
September 2023 - Present

Privacy Protector: Defending GNNs Against Inference Attacks

  • Explored existing model compression methods as a defense against Inference Attacks on GNN.
  • Developing a new training procedure leveraging weights pruning and knowledge distillation.
September 2023 - Present