A Mask Optimization Method for Automated Vulnerability Repair

In recent years, automated vulnerability repair methods, particularly those based on Transformer models, have been increasingly applied to address software security challenges. These methods often employ masking techniques to focus on attack-prone code areas. Inspired by Vision Transformers in object detection, this paper introduces a mask optimization solution. It integrates masks into the decoder’s cross-attention mechanism, allowing the model to focus on key vulnerability segments. Additionally, the mask is applied to the encoder’s self-attention, enhancing feature representation. The proposed method demonstrates improvements over approaches such as VulRepair and VQM.