litreviewslides.qmd

---
title: LLM Literature Review
author: Devin Ersoy
institute: Purdue University
date: 2024-01-17
format: revealjs
highlight-style: github
slide-number: c/t
---

## Opening A Pandora's Box: Things You Should Know in the Era of Custom GPTs
> Provides an analysis of security and privacy issues resulting from *custom* GPT*s*

<footer id="fn1" style="font-size:20px; color: #666; padding-top: 10px;">
  <a href="[https://arxiv.org/abs/2401.00905">https://arxiv.org/abs/2401.00905</a>
</footer>

## Privacy in Large Language Models: Attacks, Defenses and Future Directions

> A balance is needed between the effects privacy of LLMs and the effectiveness of the LLM resulting from the privacy enhancement.

<footer id="fn1" style="font-size:20px; color: #666; padding-top: 10px;">
  <a href="https://doi.org/10.48550/ARXIV.2310.10383">https://doi.org/10.48550/ARXIV.2310.10383</a>
</footer>

## Not What You've Signed Up For: Compromising Real-World LLM-Integrated Applications with Indirect Prompt Injection
> Adversaries can control an LLM without direct access through *indirect injection*.
- ex) Polluting the training data of an LLM

<footer id="fn1" style="font-size:20px; color: #666; padding-top: 10px;">
  <a href="https://dl.acm.org/doi/10.1145/3605764.3623985">https://dl.acm.org/doi/10.1145/3605764.3623985</a>
</footer>

## Threat due to *trust*
>
> "Models can currently act as a vulnerable, easy-to-manipulate, intermediate layer between users and information, which users might nevertheless overrely on. I.e., the model’s functionality itself can be attacked." (Abdelnabi et al., 2023, p. 5)

<footer id="fn1" style="font-size:20px; color: #666; padding-top: 10px;">
  <a href="https://dl.acm.org/doi/10.1145/3605764.3623985">https://dl.acm.org/doi/10.1145/3605764.3623985</a>
</footer>

:::notes
It is quite similar to the Malla paper
:::

## Towards Trustworthy AI Software Development Assistance
> Outlines a means to develop *trustworthy* AI assistants that provide secure and high quality code

<footer id="fn1" style="font-size:20px; color: #666; padding-top: 10px;">
  <a href="https://arxiv.org/abs/2312.09126">https://arxiv.org/abs/2312.09126</a>
</footer>

:::notes
Related to security since it is regarding *secure* code.
:::

## How Johnny Can Persuade LLMs to Jailbreak Them: Rethinking Persuasion to Challenge AI Safety by Humanizing LLMs
> Provides an analysis in how to persuade LLMs to jailbreak them through human-like communication

<footer id="fn1" style="font-size:20px; color: #666; padding-top: 10px;">
  <a href="https://arxiv.org/abs/2401.06373">https://arxiv.org/abs/2401.06373</a>
</footer>

:::notes
Explores intersection between everyday language interaction and AI safety.
:::

## A Novel Evaluation Framework for Assessing Resilience Against Prompt Injection Attacks in Large Language Models
> Provides a means to *quantify* resilience of applications against prompt injection attacks

<footer id="fn1" style="font-size:20px; color: #666; padding-top: 10px;">
  <a href="https://arxiv.org/abs/2401.00991">https://arxiv.org/abs/2401.00991</a>
</footer>

## Benchmarking and Defending Against Indirect Prompt Injection Attacks on Large Language Models
> Provides a *benchmark* for the robustness of LLMs against indirect prompt injection attacks

<footer id="fn1" style="font-size:20px; color: #666; padding-top: 10px;">
  <a href="https://arxiv.org/abs/2312.14197">https://arxiv.org/abs/2312.14197</a>
</footer>

## Can Large Language Models Identify And Reason About Security Vulnerabilities? Not Yet
> Identifies *how capable* an LLM is with regard identifying security vulnerabilities.

<footer id="fn1" style="font-size:20px; color: #666; padding-top: 10px;">
  <a href="https://arxiv.org/abs/2312.12575">https://arxiv.org/abs/2312.12575</a>
</footer>

:::notes
This one is quite similar to the LLM Software Security paper
:::