[2R4-OS-12] 機械学習品質評価・向上技術

[2R4-OS-12-01] Implementation and Evaluation of Adversarial Examples Detection Techniques Based on Network Invariants

〇Keiichi Nishida¹ (1. National Institute of Advanced Industrial Science and Technology)

Keywords:Adversarial Examples, Value Invariants

Techniques for detecting Adversarial Examples applied to the input data to a neural network include metric-based approaches, denoting approaches, prediction inconsistency-based approaches, network invariant checking approaches, etc. The most important of these approaches are the metric-based approaches. This presentation describes an implementation method for practical use of the network invariant checking approach (NIC method), which has been reported to have the highest detection rate among these approaches, and reports that it was actually able to detect adversarial perturbations with a high detection rate. The results of a discussion on the reasons for the high detection rate of the NIC method are also reported.