Progressive Prototype Alignment with Entropy Regularization for Cross-Project Software Vulnerability Detection

Cross-project software vulnerability detection must cope with pronounced domain shift and severe class imbalance, while the target project is typically unlabeled. Existing unsupervised domain adaptation techniques either focus on marginal alignment and overlook class-conditional mismatch, or depend on noisy pseudolabels, which can induce negative transfer in imbalanced settings. To address these challenges we propose DAP2ER, a progressive domain adaptation framework that couples adversarial domain confusion with entropy regularization and prototype-guided high-confidence pseudolabel optimization. Specifically, DAP2ER constructs source class prototypes, selects reliable target samples via confidence-aware pseudolabeling, and performs class-conditional alignment by pulling target features toward the corresponding prototypes. A progressive weighting schedule gradually increases the strength of domain and self-training objectives, stabilizing optimization in early epochs. Experiments on two real-world vulnerability datasets demonstrate that DAP2ER consistently outperforms strong baselines, improving the F1-score by up to 21 percentage points and achieving substantial gains in AUC for bidirectional transfer.