Adaptive Multi-level Code Graph Pruning with Vulnerability-Sensitive Retention for Robust Vulnerability Detection

It is a difficult challenge to identify software vulnerabilities in large real-world codebases correctly. It derives through noisy code charts, these far-flung semantic connections, and the way existing graph simplification techniques tend to remove nodes that are important to security. We had developed AMGP-VSR to overcome this. It is a novel system of adaptable multi-tier pruning of code graphs, and it aims at maintaining vulnerability-relevant components. The strategy draws in definite indications of the security world. Of particular importance is taint propagation. The dereference of pointers is important as well. There are functions of array indexing and bounds checking. It is rounded off by memory-API interactions. These recommend a two-score system of node importance. One of them is the score of structural importance referred to as SIS. The other is the vulnerability-sensitive score (VSS). Then there is a four-step pruning process which proceeds step by step. It eliminates up to 90 percent of the non-critical nodes. In the meantime, it retains the semantics of root-cause vulnerabilities. Subsequently, the graph is coded on the basis of a combination of GNN and graph transformer configuration. This captures local dependencies and the long-range dependencies. We did numerous tests on normal benchmark datasets. AMGP-VSR samples much better than the current systems in detecting vulnerabilities. It hit an F1-score of 0.84. That makes it 8.7 percent better than ANGEL. It is 18.3 percent superior to AMPLE. Reveal falls 25.3 percent behind. Devign is even more at 35.4 percent lower. Precision came in at 0.86. Recall was 0.82. Those figures indicate an actual reduction in false positives and false negatives. At runtime AMGP-VSR proved the fastest. The time of inference was reduced to 6.1 s. It is 32.7 percent faster than ANGEL. The actual power is revealed on huge graphs. The accuracy of detection increased by 13.6 percent at ANGEL there. It is superior in managing the noise accumulation. This was supported by ablation tests. The elimination of VSR reduced F1 by 6 percent. It was pruned off 10 percent by multi-level pruning. The Transformer backbone gone amounted to 15 percent loss. Each piece adds a lot. In general, these results make AMGP-VSR a viable and scalable vulnerability detector. It retains a great deal of semantic detail. The cost of computing reduces as well. A combination of node keeping that is security conscious and graph trimming based on adaptability opens opportunities into the future of better static analysis.