MEASER: Malware embedding attacks on open-source LLMs
Open-source large language models (LLMs) have demonstrated considerable dominance over proprietary LLMs in resolving neural processing tasks, thanks to the collaborative and sharing nature. Although full access to source codes, model parameters, and training data lays the groundwork for transparency, we argue that such a full-access manner is vulnerable to MEAs, and their ill-effects are not fully understood. In this paper, we conduct a systematic formalization for MEAs on open-source LLMs by enumerating all possible threat models associated with adversary objectives, knowledge, and capabilities. Therein, the threat posed by adversaries with internal knowledge, who inject payloads and triggers during the model sharing phase, is of practical interest. We go even further and propose the first MEA against open-source LLMs, dubbed MEASER, which wields impacts through identifying targeted parameters, embedding payloads, injecting triggers, and executing payloads sequentially. Particularly, MEASER enhances the attack robustness against quantization and parameter-efficient fine-tuning (PEFT) by employing the Magnitude-Adaptive Relative Quantization Index Modulation (MAR-QIM) mechanism, synergized with LDPC codes and spread spectrum modulation. In addition, to achieve stealthiness, MEASER devises the performance-aware importance metric to identify targeted parameters with the least degradation of model performance. Extensive experiments on four popular open-source LLMs show that the stealth rate of MEASER outperforms existing MEAs (for general DNNs) significantly, while consistently achieving a 0 bit error rate (BER) in all settings. Moreover, MEASER also maintains superior stealthiness on quantized models. We appeal for investigations on countermeasures against MEASER in view of the significant attack effectiveness.