Overview
Explore a detailed analysis of privacy backdoors in pretrained machine learning models through this comprehensive video lecture. Delve into the potential risks of fine-tuning downloaded models and learn about a method that allows attackers to fully compromise the privacy of fine-tuning data. Examine the core concept of single-use data traps, investigate how backdoors can be implemented in transformer models, and discover additional numerical techniques. Gain insights into experimental results and conclusions drawn from this research. Understand the implications of this supply chain attack on machine learning privacy and its impact on models trained with differential privacy.
Syllabus
- Intro & Overview
-Core idea: single-use data traps
- Backdoors in transformer models
- Additional numerical tricks
- Experimental results & conclusion
Taught by
Yannic Kilcher