• The University of Tokyo / RIKEN
  • tsuzukums.k.u-tokyo.ac.jp

About Me

Profile.

Blog

My personal blog.

CS/ML Blog

Topics related to CS/ML.


Welcome To My Site

This is a personal site of Yusuke Tsuzuku.
Focus of this site is Computer Science and Machine Learning.
This site is under construction.

News

Recent Projects

Black-box Universal Adversarial Perturbations

Dec, 2017 Web Design

Most sensitive directions of convolutional layers are some specific frequency in Fourier basis functions. We exploit this fact to attack deep convolutional networks. Experimentally, adding just single Fourier basis function to input image significantly degraded accuracy of networks. Thus, we can create universal adversarial perturbations by just selecting a frequency, which enables us an efficient black-box attack.


Read More

Lipschitz-Margin Training

Dec, 2017 Web Design

High sensitivity of neural networks against malicious perturbations on inputs causes security concerns. The evaluation of defense against such perturbations are notoriously difficult. Thus, we aim to obtain certifiable robustness. However, prior certification work requires strong assumptions on network structures and massive computational costs, and thus, their applications are limited. We present a computationally efficient calculation technique that lower-bounds the size of adversarial perturbations.


Read More



Featured From Blog

2019年1月1日

Dec, 2017 Web Design

Read More

海外Ph.D.もいい気がしてきた

Dec, 2017 Web Design

そろそろ進路について真面目に考える必要がある

Read More

留学費用

Dec, 2017 Web Design

支援

Read More