Beyond Statistical Correlation: Enforcing Crash Consistency and Semantic Distillation for Accurate Root Cause Analysis

Root cause analysis (RCA) plays a crucial role in diagnosing and fixing software vulnerabilities. Among various RCA approaches, statistical approaches have shown promise due to their generality. However, existing statistical RCA frameworks often suffer from low crash consistency and noise from unrelated predicates, limiting their effectiveness. To address these limitations, we propose ASRCA, a novel statistical RCA framework that integrates a two-stage filtering strategy and backward data-flow analysis to enforce crash consistency and reduce semantic noise, thereby improving accuracy. We implement ASRCA atop the AURORA framework and evaluate it on five real-world vulnerabilities. Experimental results show that ASRCA achieves $\mathbf{1 0 0 \%}$ crash consistency and successfully identifies the root cause in all cases, ranking it within the top 20 for every vulnerability. Compared to AURORA, ASRCA significantly improves both the accuracy and efficiency.