Although deep neural networks have been shown to be successful in various computer vision tasks such as image classification and biometric recognition, robustness and reliability are always the concerns of using deep learning models. On the one hand, degraded images and videos aggravate the performance of biometric recognition. On the other hand, if the deep neural networks are under adversarial attacks, the networks can be broken completely. Motivated by the vulnerability of deep neural networks, I analyze and develop robust biometric recognition including image restoration and adversarial defense algorithms towards a vision of robust machine learning in computer vision.

In this talk, I study two types of degradation making deep neural networks vulnerable. The first part of the talk focuses on face recognition at long range, whose performance is severely degraded by atmospheric turbulence. The theme is on improving the performance and robustness of various tasks in face recognitionsystems such as face detection, facial keypoints localization, feature extraction, and image restoration. The second part focuses on defending adversarial attacks in the images classification and face recognition tasks. The theme is on exploring adversarial defense methods that can achieve good performance in standard accuracy, robustness to adversarial attacks with known threat models, and good generalization to other unseen attacks.

Speaker: Dr Chun Pong LAU
Date: 24 April 2023 (Monday)
Time: 9:00am – 10:00am
Poster: Click here