Spinning Language Models: Risks of Propaganda-as-a-Service and Countermeasures

Explore the potential risks and countermeasures associated with using language models for propaganda in this 17-minute IEEE conference talk. Delve into the concept of Propaganda-as-a-Service, examining the differences between classification and sequence-to-sequence models. Gain insights into the intuition behind spinning language models and learn about the creation of input for meta-task models. Analyze the changes in output distribution and discover defensive strategies to mitigate these risks. Presented by Eugene Bagdasaryan and Vitaly Shmatikov from Cornell Tech, this talk offers a comprehensive overview of the challenges posed by manipulated language models in the context of propaganda dissemination.

Intro
What is Propaganda?
Classification vs Sequence-to-sequence
Spin Intuition in Language Models
Create Input for Meta-task Model
Change in output distribution
Defense