Securing IMF from Intent Injection Using BERT-Based Classifiers

Intent-based systems empower autonomous networks to operate based on high-level declarative goals, offering flexibility, scalability, and reduced operational complexity. While the security implications of such systems are recognized, intent injection attacks, which adversaries manipulate syntactically valid but semantically malicious intents, have remained largely theoretical due to the lack of standardized intent samples for empirical evaluation. This paper presents a semantic-level injection detection framework that employs a fine-tuned BERT classifier to transform each intent into a natural language representation. This enables semantic analysis of the intent's content and classifies it as either benign or injected. To bridge the gap of unavailable real-world data, we synthesize training samples guided by benign and adversarial intent policies. For evaluation, we use intent samples from the recently released ETSI 3GPP TS 28.312 specification, creating variants by injecting code into their YAML structures to preserve structural validity while altering semantic meaning. Experimental results show that the proposed method effectively detects injected intents, achieving an AUC of 0.98 and an F1-score of 0.97, significantly outperforming existing injection detection methods such as PromptGuard and VulBERT, which prove ineffective against these attacks. These results underscore the necessity of developing tailored detection mechanisms for intent-based systems and highlight the potential of semantics-aware approaches to address this unknown known security threats.