Enhancing the security of edge-AI runtime environments: a fine-tuning method based on large language models

Leveraging the convenience of container technology, rapidly deploying complex computing environments and systems across different architectures, such as those found in edge AI scenarios, has become feasible. The security of the runtime environment is critical in the cross-architecture deployment of container-based Edge-AI models, as it directly affects model stability and privacy. Consequently, substantial research efforts have been dedicated to developing learning-based container escape detectors to ensure the security of the runtime environment. However, the efficacy of current detection methods is significantly dependent on the size and quality of training samples, which are constrained by the computing resources available in edge environments. Drawing inspiration from the exceptional performance of large language models in natural language generation and understanding tasks, this paper proposes a lightweight joint fine-tuning strategy based on Prefix and LoRA. By fine-tuning ChatGPT and ChatGLM, we aim to automatically generate high-quality container escape data samples to enhance existing detectors. Comprehensive experimental evaluations using three state-of-the-art container escape detectors reveal that the samples generated through this approach can substantially improve the performance and robustness of these detectors.