Impacts of Artefacts and Adversarial Attacks in Deep Learning based Action Recognition

Abstract

Current state-of-the-art deep learning-based models for human action recognition achieve impressive accuracy on benchmark datasets. However, the fact that those models are trained and tested on “clean” and high-quality input data raises a concern about their reliability under transmission artefacts and adversarial perturbations. In this work, we conduct for the first time an evaluation of the impacts of artefacts and adversarial attacks in deep learning-based human action recognition. Findings from this evaluation provide insights into the behaviors of action recognition under hostile conditions of best-effort networks.