Widening the Gap: Exploiting LLM Quantization via Outlier Injection

Researchers have discovered a vulnerability in large language models (LLMs) that allows attackers to inject malicious behavior into quantized models. The attack exploits the process of quantization, which reduces the precision of model weights to decrease memory usage. By carefully crafting input data, an adversary can create a model that appears benign in its full-precision form but exhibits malicious behavior when quantized. This vulnerability poses a significant risk, as quantized models are increasingly used in resource-constrained environments. The attack is particularly effective against models that use simple quantization methods, highlighting the need for more robust quantization schemes. The findings demonstrate that quantization can introduce significant security risks, and practitioners must be aware of these risks when deploying quantized models¹. This matters to security practitioners because it highlights the need for careful evaluation of quantized models to ensure they do not introduce unforeseen vulnerabilities.