核心概念
提案されたTSFoolアプローチは、RNNベースの時系列分類における高度に認識しにくい敵対的な時系列を効率的に作成し、既存の手法を効果、効率、認識性の観点で大幅に上回っています。
統計
NNs are vulnerable to adversarial attacks [52], which means imperceptible perturbations added to the input can cause the output to change significantly [28].
Recent years have witnessed the success of recurrent neural network (RNN) models in time series classification (TSC).
Experiments on 11 UCR and UEA datasets showcase that TSFool significantly outperforms six white-box and three black-box benchmark attacks in terms of effectiveness, efficiency and imperceptibility.
引用
"By exploring the visual sensitivity of time series data, for the first time, we point out the bias of the conventional local optimization objective of adversarial attack."
"We propose a novel optimization objective named “Camouflage Coefficient" to enhance the global imperceptibility of adversarial samples."