presentation on the article "Deep leakage from the gradients"

Created using ChatSlide

In this lecture, we will explore the critical aspects and implications of gradient privacy in machine learning (ML). Initially assumed safe, gradients have been disproved as secure through recent studies like Deep Leakage from Gradients (DLG), revealing significant privacy concerns. We will delve into related work highlighting previous research limitations, our novel methodology for data reconstruction, and experimental results across vision and NLP tasks. Threat analyses will illuminate...