Vib-ner enables out-of-vocabulary recognition in cybersecurity threat

VIB-NER, a new AI model unveiled by researchers in Xinjiang, China, is designed to better recognize threat terms that have never appeared before in cybersecurity reports. Built by a team at the National Security Research Institute at Shihezi University and described in a study published in the journal Cybersecurity, the system combines a variational information bottleneck approach with a mutual information-based loss function. It posts F1 scores of around 79%, recall of 77% and precision of 80% on cybersecurity threat intelligence entity extraction tasks—about a 4% to 8% improvement over mainstream E-NER models. The researchers also report halving the training time per batch. The work targets weaknesses in named entity recognition caused by out-of-vocabulary hashes, vulnerability IDs, malware names and hacker group labels.