Adversarial AI Attacks, Mitigations, and Defense Strategies

13

LLM Foundations for Adversarial AI

In the previous chapter, we started looking at Generative AI and the profound challenges it brings in its ability to generate outputs that can be used adversarially.

Despite the different outputs, our first exploration of Generative AI still accepted inputs conventionally – that is, encoded images for GANs. In this chapter, we will look at large language models (LLMs) and their use of prompts, free-text inputs, mixing content, and instructions for the model. LLMs are a field of their own, with a very different development workflow. We will look at prompts in the context of the LLM revolution sparked by ChatGPT and the slight paradigm shift toward accessing external hosts via APIs rather than the model directly. ...