计算机科学
有效载荷(计算)
嵌入式系统
操作系统
编码(内存)
吞吐量
突变
算法设计
编码(集合论)
计算机网络
电子邮件
实时计算
SQL语言
并行计算
作者
Yujia Shi,Ruonan Li,Jinlong Li,LiChen Liu,Ye Wang,Hengyuan Li,Yujia Xiao,Zhaoquan Gu,Jie Liu
标识
DOI:10.1109/tdsc.2026.3731062
摘要
Web Application Firewalls (WAFs) are widely deployed as a core defense for web applications. Evaluating their defensive capabilities against high-impact threats such as SQL injection (SQLi) relies on the quality of test payloads. That quality is largely determined by the payload generation method. Most existing generators apply token-level perturbations under fixed, handcrafted rules, which yields restricted syntactic variation and shallow compositional structures, limiting their ability to expose WAF blind spots. To address these challenges, we propose Vast-SQLi, a constraint-guided framework that generates SQLi test payloads via structure-aware rewrites. Payloads are encoded as customized abstract syntax trees (ASTs), allowing structural transformations beyond token-level edits. These ASTs are used to fine-tune large language models on transformation traces that capture domain-specific editing strategies. Reinforcement learning then employs constraint-guided objectives to generate valid payloads without reliance on handcrafted obfuscation rules. Evaluations on eight representative WAFs show Vast-SQLi improves bypass success by 27.43% over raw inputs and surpasses state-of-the-art methods. The results are shared with WAF developers to enhance defenses and guide future evaluation practices.
科研通智能强力驱动
Strongly Powered by AbleSci AI