Backdoor Learning: Recent Advances and Future Trends


In conjunction with ICCV 2023

October 3th, 2023 (13:30 - 17:20 CEST)

Location: Room W08, The Paris Convention Center in Paris, France

ICCV 2023 Tutorial on "Backdoor Learning: Recent Advances and Future Trends"

Backdoor learning is an emerging and crucial field of research focused on investigating the security of machine learning systems, specifically computer vision systems, during the training phase. It has been demonstrated that an adversary could manipulate the training process to insert a backdoor into the trained model, such that the backdoored model will perform well on benign images while producing an adversary-specified prediction on images that has been tampered with. This tutorial aims to provide a comprehensive and detailed introduction to the field of backdoor learning, covering a wide range of important and interesting topics. We start by presenting basic definitions and taxonomies that are essential to understand the concept of backdoor learning. Then, we dive into the current progress of the field by presenting various existing attacks and defenses highlighting the seriousness of the threats and challenges faced by machine learning systems during their train- ing phase. After that, we will discuss the latest benchmark that has been developed for backdoor learning. To conclude the tutorial, we will discuss the real-world applications of backdoor learning and the challenges and future trends in this exciting research area.

Schedule

  • Part I: Definitions and taxonomies of backdoor learning [Baoyuan Wu, 20 mins, 13:30 - 13:50] [slides]
  • Part II: Backdoor attacks [Baoyuan Wu, 60 mins, 13:50 - 14:50] [slides]
    • Data poisoning-based backdoor attacks
    • Training-controllable backdoor attacks
  • Coffee Break [20 mins, 14:50 - 15:10]:
  • Part III: Backdoor defenses [Baoyuan Wu & Mingli Zhu & Shaokui Wei, 60 mins, 15:10 - 16:10] [slides]
    • Pre-processing stage backdoor defenses
    • In-training stage backdoor defenses
    • Post-training stage backdoor defenses
    • Inference stage backdoor defenses
  • Coffee Break [20 mins, 16:10 - 16:30]:
  • Part IV: The latest comprehensive benchmark for backdoor learning [Baoyuan Wu, 20 mins, 16:30 - 16:50] [slides]
  • Part V: Applications and use cases of backdoor learning [Hasan Hammoud, 30 mins, 16:50 - 17:20] [slides]

Materials

Organizers

Baoyuan Wu

Associate Professor at CUHK-SZ

Bernard Ghanem

Professor at KAUST

Hasan Abed Al Kader Hammoud

PhD student at KAUST

Contacts

Contact the Organizing Committee: wubaoyuan@cuhk.edu.cn, bernard.ghanem@kaust.edu.sa, hasanabedalkader.hammoud@kaust.edu.sa