Capturing the Effects of Quantization on Trojans in Code LLMs

TL;DR

This paper investigates how model quantization affects the vulnerability of code-generating large language models to trojan attacks, revealing that quantization can both mitigate and have limited impact on such risks depending on the model.

Contribution

It introduces a new metric for trojan signals and analyzes the differential effects of quantization on two large language models in the context of code generation.

Findings

Quantization at 4-bit improves CodeLlama's robustness against trojan attacks.

Quantization has minimal impact on Llama-2's behavior and vulnerability.

A new metric for measuring trojan signals in models is proposed.

Abstract

Large language models of code exhibit high capability in performing diverse software engineering tasks, such as code translation, defect detection, text-to-code generation, and code summarization. While their ability to enhance developer productivity has spurred widespread use, these models have also seen substantial growth in size, often reaching billions of parameters. This scale demands efficient memory resource usage, prompting practitioners to use optimization techniques such as model quantization. Quantization uses smaller bit representations for the model parameters, reducing the precision of the weights. In this work, we investigate the impact of quantization on the risk of data poisoning attacks on these models, specifically examining whether it mitigates or exacerbates such vulnerabilities. We focus on two large language models, Meta's Llama-2-7b and CodeLlama-7b, applied to an SQL code generation task. Additionally, we introduce a new metric for measuring trojan signals in compromised models. We find that quantization has differing effects on code-generating LLMs: while reducing precision does not significantly alter Llama-2's behavior, it boosts performance and reduces attack success rates in CodeLlama, particularly at 4-bit precision.