Exploring Transformers for Multi-Label Classification of Java Vulnerabilities

Deep learning (DL) techniques have demonstrated potential in reasoning complex patterns of vulnerable code from high-level abstractions. Recent advancements in the area, such as the introduction of transformer-based models, like BERT, help overcome the problem of the available vulnerability detection datasets being too small to enable most DL models to capture all relevant patterns. They mitigate the challenge by leveraging knowledge from a general domain to solve problems in specific domains. In this paper, we explore different BERT-based models for multi-label classification of vulnerabilities in Java on a synthetic dataset. The models yield up to 99% in accuracy and 94% in f1-score. We remove biases in the training dataset and observe drops of up to 13% of the f1-score. We further assess the generalizability of the models on realistic samples and notice that one model, in particular, predicted unknown vulnerabilities with an f1-score of nearly 85%.