Adversarial Machine Learning
Oxide and Friends
00:00
Exploring Adversarial Tactics on Language Models
The chapter delves into the transition from attacking image recognition models to exploring adversarial tactics on language models. It discusses the applications of multi-modal models, challenges of adapting adversarial techniques to language models, and the vulnerability of models to subtle perturbations. The conversation also touches on training anti-models, manipulating models to produce specific responses in chat models, and the strategy of optimizing inputs with the word 'sure' to prompt negative or harmful responses.
Transcript
Play full episode
