One property of adversarial examples is transferability: adversarial examples for one model can also mislead other models.
We explore how to break the transferability for adversarial examples, i.e, to train a complementary model given a reference model (adversarial examples cannot transfer across them).
The problem is formalized as a non-convex min-max game problem.
We propose a convex approximation to effectively solve the problem.
Regexs generation from NL sentences
We focus on automatically generating regular expressions from natural language specifications.
Existing approaches reduce the target problem to a general machine translation task.
We find a huge drop of performance when applying existing approaches on real-world data vs synthetic data.
We propose to utilize string examples of regexs to define the semantic correctness, and use reinforcement learning to maximize the semantic correctness of regexs predicted by the model.
User Identity Linkage
User Identity Linkage refers to match identical users across two social networks.
In this project, we propose an unsupervised framework based on co-training to address this problem.
We develop an attribute-based model and a relationship-based model. The models will be enhanced in iterative way via co-training.
We also propose to employ sequence-to-sequence learning to measure the similarity of attributes, which is an effective implementation of attribute-based model.
DualFocus is a tool for image refocusing. DualFocus takes two images with different focal lengths as input, and outputs an all-in-focus image.
For each image, we solve a deconvolution problem with constrains based on prior characteristics of natural in-focus images.
We then use Alternating Direction Method of Multipliers (ADMM) to minimize the differences between the deconvolution results for two input images.
DualFocus does not require modification of cameras (i.e., coded aperture), and outperformances single-image-based refocusing approaches.
ObjRender is a tool to render 3D synthetic objects in images. We use a mirror ball as a light probe to obtain environment mapping.
We take several low dynamic range (LDR) images of the mirror ball to recover a high dynamic range (HDR) image. We build the environment lighting box by mapping the HDR image in sphere domain to the equirectangular domain. We annotate necessary geometry models such as tables in the scene, and render the synthetic objects on those models.
2017.8 - present, University of Illinois at Urbana-Champaign
M.S. in Computer Science
2013.9 - 2017.7, Peking University
B.S. in Computer Science and Technology
2017.8 - present, Research Assistant, University of Illinois at Urbana-Champaign
Advisor: Prof. Tao Xie
2018.5 - 2018.8, Visiting Research Assistant, MIT Media Lab
Advisor: Prof. Fadel Adib
2016.7 - 2017.5, Research Intern, Microsoft Research Asia
Mentors: Dr. Zaiqing Nie and Dr. Yong Cao
2015.11 - 2016.5, Research Assistant, Peking University
Advisor: Prof. Yingfei Xiong