Privacy attacks and LLMs
In Chapters 8 and 9, we discussed privacy attacks on AI that steal models and sensitive data or infer sensitive data in detail. Our discussion was in the context of predictive AI, but recent research has validated that these attacks also apply to LLMs.
Two good research papers provide comprehensive surveys of related research:
- Privacy in Large Language Models: Attacks, Defenses and Future Directions by Li, Chen, and others, published in 2023 at https://arxiv.org/abs/2310.10383
- A survey on Large Language Model (LLM) security and privacy: The Good, The Bad, and The Ugly by Yao, Duan, and others, published in 2024 at https://www.sciencedirect.com/science/article/pii/S266729522400014X
As with all our previous discussions on LLMs, what makes adversarial attacks on LLMs is affected by the following:
- Size: This affects both the model size and the web-scale size of data used to train FMs, which makes it highly challenging to ensure data...
