| 研究実績の概要 |
Highlights of the research outcomes for the final year of the project included:
1. Acceptance of a refereed international conference paper at ICLR 2018 (presented on 2018/5/1). Our paper deals with the problem of detecting adversarial examples, which are carefully crafted instances that can mislead deep neural networks (DNNs) to make errors during prediction. We show that adversarial examples can be characterized in terms of the unusually high local intrinsic dimensionality (LID) within the data space surrounding them, according to the model of LID developed in this project. Our paper was one of only 23 papers (out of 935) accepted for full oral presentation (acceptance rate: 2.5%).
2. Presentation of a refereed international conference paper at WIFS 2017, in which a theoretical analysis is given showing that the vulnerability of classification to adversarial perturbation increases as the local intrinsic dimensionality rises.
3. Presentation of 3 refereed international conference papers (at SISAP 2017). Two of these papers together lay out the full theoretical foundations of the LID model, its connection to similarity search and extreme value theory, and an extension to the multivariate setting. The third paper showed the power of the LID model for the practical guiding the selection of local features suitable for the formation of high-quality similarity graphs.
4. A refereed international journal publication which uses the theory of local intrinsic dimensionality to guide the heuristic termination of content-based similarity search processes involving multiple query objects.
|
