Investigating the significance of adversarial attacks and their relation to interpretability for radar-based human activity recognition systems
作者:
Highlights:
• We analyze adversarial threats for human action recognition using radar data.
• We expose the vulnerability of radar-based models to adversarial inputs.
• We present adversarial padding attack to be used on data with a temporal dimension.
• We show the possibility of changing prediction without perturbing activity frames
• We aim at linking adversarial vulnerability and prediction interpretability.
摘要
•We analyze adversarial threats for human action recognition using radar data.•We expose the vulnerability of radar-based models to adversarial inputs.•We present adversarial padding attack to be used on data with a temporal dimension.•We show the possibility of changing prediction without perturbing activity frames•We aim at linking adversarial vulnerability and prediction interpretability.
论文关键词:Radar data,Activity recognition,Adversarial examples,Neural network interpretability,Deep convolutional neural networks
论文评审过程:Received 1 October 2019, Revised 14 August 2020, Accepted 16 September 2020, Available online 23 September 2020, Version of Record 12 October 2020.
论文官网地址:https://doi.org/10.1016/j.cviu.2020.103111