Course for: Department of Computer Science, Aalborg University, Denmark
Course time and location: Tu & Th 1:00 PM - 3:00 PM (CEST) Online
Instructor: Jiliang Tang, Han Xu
In recent years, deep learning is shown to be a key technique that leads to remarkable breakthroughs of many AI applications. In particular, trustworthy AI becomes an emerging topic to promote the trust of AI and enable a better understanding of the pros and cons of deep learning systems. In particular, the idea of adversarial attacks and defenses is crucial for enhancing the learning reliability, ensuring trustworthy decision making, and improving the understanding of machine learning methods. Algorithm fairness stress the models should equally treat different individual for model prediction. Besides, recent advances in Generative AIs, such as large language models (LLMs) and diffusion models (DMs), also bring new challenges about the exploring and overcoming the trustworthy problems in Generative Models. This course aims to equip students with a comprehensive understanding of the key concepts and methodologies for studying the potential and related problems in real-world applications.
Lecture 1 (Oct 10): Introduction to Trustworthy AI and adversarial robustness. [slides]
1. Understand different aspects of trustworthy AI problems in various domains.
2. Algorithms on adversarial attacks.
Lecture 2 (Oct 12): Defenses against adversarial attacks. [slides]
1. Adversarial training.
2. Certified Defenses against adversarial examples.
Lecture 3: Data poisoning attacks (Oct 17). [slides]
1. Data poisoning attacks in linear models.
2. Data poisoning attacks in deep learning models: targeted attacks and backdoor attacks.
3. Unlearnable examples.
Lecture 4: Machine learning fairness (Oct 19). [slides]
1. Algorithms to achieve fairness in linear models.
2. Algorithms to achieve fairness in deep learning models.
Lecture 5: Trustworthy AI in generative models (Oct 24). [slides]
1. Trustworthy risks and countermeasures in large language models.
2. Trustworthy risks and countermeasures in diffusion models.