Automated vulnerability detection has become increasingly important. Many existing methods utilize deep learning models to obtain code representations for vulnerability detection. However, these approaches predominantly capture the overall semantics of the code rather than its intrinsic vulnerability-specific semantics. To address this issue, we propose CLeVeR, the first approach that leverages contrastive learning to generate precise vulnerability code representations under the supervision of vulnerability descriptions. Specifically, we introduce an Adapter, a Representation Refinement module, and a Description Simulator to mitigate the challenges of semantic misalignment and imbalance between code and descriptions, and input data inconsistency between pre-training and fine-tuning stages, respectively. For vulnerability detection and classification tasks, CLeVeR achieves F1 scores of 72.82% (real-world dataset) and 80.34%, outperforming state-of-the-art methods (SOTAs) by 11.85% and 13.61%. Additionally, CLeVeR also outperforms SOTAs in zero-shot inference, demonstrating the transferability of its generated vulnerability code representations.

Transformer-based models are not efficient in processing long sequences due to the quadratic space and time complexity of the self-attention modules. To address this limitation, Linformer and Informer reduce the quadratic complexity to linear (modulo logarithmic factors) via low-dimensional projection and row selection, respectively. These two models are intrinsically connected, and to understand their connection we introduce a theoretical framework of matrix sketching. Based on the theoretical analysis, we propose Skeinformer to accelerate self-attention and further improve the accuracy of matrix approximation to self-attention with column sampling, adaptive row normalization and pilot sampling reutilization. Experiments on the Long Range Arena benchmark demonstrate that our methods outperform alternatives with a consistently smaller time/space footprint.