Scholarlink

<p>Traditional web fuzzing frequently exhibits limited efficiency and insufficient coverage, particularly within complex environments such as WAF-protected interfaces and deep parameter injection scenarios. To mitigate these issues, we propose MHFuzzer, a framework underpinned by the HCTNet hybrid architecture. HCTNet synthesizes multi-scale causal convolutions, dual-layer unidirectional LSTMs, and causality-constrained Transformers via cross-attention to extract deep semantic patterns. Additionally, MHFuzzer integrates mean shift clustering, hierarchical scheduling, and reinforcement learning to establish an adaptive optimization loop. Experiments on CSIC-2010 and OWASP-Benchmark validate superior path exploration capabilities. Specifically, in WAF-bypass scenarios, the proposed method increases the quantity of effective payloads by over 67% compared to the RL-Agent baseline, while achieving faster convergence. By employing a cascaded dependency model and a probability correction dictionary, the model attains a 90.3% generation accuracy on CSIC-2010. Furthermore, payload accuracy reaches 85.3% in high-difficulty WAF evasion contexts, outperforming Transformer and GPT-Fuzzer models while minimizing false positives. Overall, MHFuzzer effectively balances coverage, detection capability, and resource efficiency.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Towards high-coverage web fuzzing: a framework integrating hybrid semantic modeling and online adaptive optimization

  • 期刊论文

  • 发表日期：

    2026年3月10日

查看全文
  • Zihao Zhou,
  • Kaibin Wei,
  • LingJie Yuan,
  • Xiannian Xie

摘要

Traditional web fuzzing frequently exhibits limited efficiency and insufficient coverage, particularly within complex environments such as WAF-protected interfaces and deep parameter injection scenarios. To mitigate these issues, we propose MHFuzzer, a framework underpinned by the HCTNet hybrid architecture. HCTNet synthesizes multi-scale causal convolutions, dual-layer unidirectional LSTMs, and causality-constrained Transformers via cross-attention to extract deep semantic patterns. Additionally, MHFuzzer integrates mean shift clustering, hierarchical scheduling, and reinforcement learning to establish an adaptive optimization loop. Experiments on CSIC-2010 and OWASP-Benchmark validate superior path exploration capabilities. Specifically, in WAF-bypass scenarios, the proposed method increases the quantity of effective payloads by over 67% compared to the RL-Agent baseline, while achieving faster convergence. By employing a cascaded dependency model and a probability correction dictionary, the model attains a 90.3% generation accuracy on CSIC-2010. Furthermore, payload accuracy reaches 85.3% in high-difficulty WAF evasion contexts, outperforming Transformer and GPT-Fuzzer models while minimizing false positives. Overall, MHFuzzer effectively balances coverage, detection capability, and resource efficiency.