Learning Internal Representations Robust against Adversarial Attacks

2021 Fiscal Year Final Research Report

• Project/Area Number: 20K19824
• Research Category: Grant-in-Aid for Early-Career Scientists
• Allocation Type: Multi-year Fund
• Review Section: Basic Section 61010:Perceptual information processing-related
• Research Institution: Kyushu University
• Principal Investigator: Vargas Danilo 九州大学, システム情報科学研究院, 准教授 (00795536)
• Project Period (FY): 2020-04-01 – 2022-03-31
• Keywords: Robust AI / Robust Machine Learning

Outline of Final Research Achievements

Here, I proposed to tackle the robustness of DNNs by evaluating and improving the internal representation learned by DNNs. Regarding the evaluation of the internal representation of DNNs, we discovered that the transferability of features links to robustness to adversarial attacks. In other words, the better the transfer of features the better the robustness to adversarial attacks. We also proposed K-spectrum which can evaluate and visualize multiple layers of DNNs together in a graph, allowing for easy inspection of how their shapes are in multi-dimensional space. Regarding the improvement of the internal representation of DNNs, we have developed as described in the proposition a GAN based system to improve the network robustness. The system outperformed the state-of-the-art and is being submitted to a journal now.
Results of this research were published in journals and proceedings, more than 13 articles in total.

Free Research Field

Artificial Intelligence

Academic Significance and Societal Importance of the Research Achievements

Critical systems such as autonomous driving and medical applications require robust machine learning algorithms. This research paves the way to better algorithms that will allow for such applications to become a reality.