Network intrusions attacks represent a significant modern security challenge, as exploitation of vulnerabilities in information systems can lead to unauthorized access to sensitive data, disruption of critical services, financial damage, a reduction in the confidentiality, integrity, and availability of any information system. Machine learning enables automated detection of network intrusions; however, such systems typically suffer from two major limitations: network data is processed \enquote{far} from the point of capture, and the user is returned only a class label (e.g., DDoS, PortScan), without any explanation of the basis for the decision.
In this work, we address the explainability of such predictions using large language models (LLMs) implemented on a smart network interface card (SNIC) directly in the network. We design and implement the FlowExplain system, which automatically generates explanations from the predictions of a simple classifier and network flow features with the help of an LLM. FlowExplain is based on a fine-tuned, smaller local model, for which specially collected training examples of network intrusions, labeled by experts, are used for fine-tuning. With FlowExplain, we show that automated, explainable network intrusion detection on local infrastructure is feasible in practice.
|