Close

Minhao Cheng

Phd Student @ UCLA CS

Download Resume

About Me

Hi, I am a Ph.D. student of Computer Science at UCLA, advised by Cho-Jui Hsieh . Prior to coming to UCLA, I was a Ph.D. student in UC Davis during 2015-2018. I received my bachelor degree in Computer Science from University of Electronic Science and Technology of China, advised by Yu Xiang. My research interests includes large-scale optimization, distributed and parallel machine learning, recommender system and robustness of machine learning models. You could reach me at mhcheng@ucla.edu.

Experience

Adversarial Attack on Interactive Dialog System

Although we have shown that the sequence-to-sequence model is not robust against adversarial attack, it is still a open question about the robustness of dialog system implemented by deep neural networks. We develop three algorithms which could attack the goal-based dialog system successfully with or without knowing the deep neural network model structure. And we could improve the model's robustness by adversarial training.

Query Efficient Hard-label Black-box Attack

It has been proved that DNNs models are vulnerable to a very small human-imperceptible perturbation. However, it is a still a challenge when we could only get hard-label instead of probability output. We develop a query efficient algorithm which could apply to industrial-strength image classifiers.

Adversarial Example for Sequence to Sequence Model

Recent research on DNNs has indicated ever-increasing concern on the robustness to adversarial examples. We are designing algorithms to generate adversarial example for sequence to sequence model which is widely used in machine translation, text summarization.

Low-rank Approximation of rankSVM in Recommendation System

New low-rank approximation method using for the recommendation system. Previously, it uses the ranksvm and its variations to train the data. Now we use the low-rank method to get better speed and accuracy dealing with the pair-wised data in ranksvm.

Line Search Method for Distributed Primal-Dual Optimization

New line search method in stochastic algorithms in large-scale distributed machine learning scenario to overcome the large Primal-Dual gap in the training periods so that we can achieve a faster convergence and a better accuracy in the distributed machine learning case.

Education

University of California Los Angeles

Sept 2018 - now

Ph.D. student in Computer Science

University of California Davis

Sept 2015 - Sept 2018

Ph.D. student in Computer Science

University of Electronic Science and Technology of China

Sept 2011 - June 2015

Bachelor of Science in Computer Science

Skills

Publications

Conference Paper

Pre-prints

Patents