Scrubbing Sensitive Information
The discussion delves into the challenges of ensuring AI models do not retain sensitive information, even after fine-tuning. Peter explains how traditional methods may leave residual knowledge within the model, while he proposes a more thorough approach to erase this information from both the final classification layer and intermediate representations. This robust method aims to enhance model safety against potential probing attacks.In this clip
From this podcast
The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence)
Localizing and Editing Knowledge in LLMs with Peter Hase - 679
Related Questions
How can we defend against adversarial attacks on machine learning models as discussed in the episode Localizing and Editing Knowledge in LLMs with Peter Hase - 679?
How can we defend against adversarial attacks on machine learning models as discussed in the episode Localizing and Editing Knowledge in LLMs with Peter Hase - 679 and the clip Adversarial Dynamics?
How can we defend against adversarial attacks on machine learning models?