Design and Validation of a Verification Platform for Enhancing Robustness of Image Object Recognition Systems
Annotatsiya
Intelligent object recognition models based on deep neural networks are not robust to small intentionally designed perturbations, so the robustness of such systems needs to be enhanced and verified. In this paper, we design and build a verification platform for improving the robustness of image-based object recognition systems, which integrates nine mainstream robustness enhancement algorithms. We conducted platform applicability validation experiments and performance comparison experiments for different types of robustness enhancement methods. The experimental results show that methods based on adversarial example detection achieve the highest recognition accuracy, and methods based on JPEG compression can effectively enhance the robustness of object recognition systems. The robustness enhancement verification platform built in this paper can effectively validate the effects of different robustness enhancement strategies on the robustness of object recognition systems.