Abstract
It is possible to retrieve data from security-related occurrences, Named Entity Recognition (NER) is essential to cybersecurity. For rich contextual text embeddings, current approaches rely on pre-trained models; nevertheless, anisotropy presents a difficulty that may impact the quality of subsequent encoding. Furthermore, current models might have trouble detecting clutter. In order to deal with these problems, we provide a unique model that combines Belief Rule Base with Contrastive Learning for Named Entry Recognition (NER) in cybersecurity, called Hybrid Contrastive Learning (HCL), which is based on deep learning. Additionally, as a BRB parameter optimization method, the Distributed Constraint Covariance Matrix Adaptation Evolution Strategy (D-CMA-ES) is proposed. This paper contributes to advancing the state of the art in NER and provides insights into building more effective, interpretable, and scalable models for cybersecurity applications. Modelling and recognising entities across a wide range of cybersecurity data is crucial for effective and efficient response to cybersecurity crises. Neural networks are being used for entity extraction in the field of cybersecurity since Named Entity Recognition (NER) was developed. BRB used to improve the detection of fixed format entities is feasible and beneficial. As an alternative to the CMA-ES technique, we suggested the D-CMA-ES algorithm, which adaptively divides data into multiple subspaces for sampling, thereby mitigating the negative impact of high-dimensional samples on training outcomes. Experiments show that NER accuracy for cybersecurity is significantly improved when HCL is combined with the D-CMA-ES algorithm.
Keywords: Belief Rules Base, Cybersecurity, D-CMA-ES, NER