Selected work

Representative Papers

ShellForge: Adversarial Co-Evolution of Webshell Generation and Multi-View Detection for Robust Webshell Defense

Jan 28, 2026

This work addresses the challenge that existing webshell detection methods struggle to keep pace with rapidly evolving obfuscated variants and suffer from high false-positive rates on benign obfuscated scripts. To overcome these limitations, the authors propose an adversarial co-evolution framework in which a generator and a multi-perspective detector iteratively train each other to continuously enhance defensive capabilities. The approach innovatively leverages de-maliciousized samples as high-quality hard negatives and integrates supervised fine-tuning with preference-based reinforcement learning to generate highly evasive webshells. The detector combines multi-view features, including semantic compression of long strings, pruned abstract syntax tree (AST) structures, and Shannon entropy. Evaluated on the FWOID benchmark, the method achieves an F1-score of 0.981, and its generated samples attain a 0.939 evasion success rate against commercial VirusTotal engines, significantly reducing false positives while improving generalization.

0 citationsRead paper

MEEA: Mere Exposure Effect-Driven Confrontational Optimization for LLM Jailbreaking

Dec 21, 2025

Existing LLM safety alignment research predominantly relies on static boundary assumptions, neglecting the behavioral dynamics induced by contextual interactions—thereby limiting robustness and generalization. Method: We propose the first multi-round adversarial evaluation framework integrating psychology’s “mere-exposure effect,” employing semantically progressive, low-toxicity prompt chains to repeatedly expose black-box models and dynamically lower their safety thresholds. Our approach combines simulated annealing optimization, joint toxicity-similarity guidance, and strategic multi-turn interaction. Results: Experiments across major models—including GPT-4, Claude-3.5, and DeepSeek-R1—demonstrate an average attack success rate improvement exceeding 20% over seven baseline methods. This work provides the first empirical evidence of historical dependence and dynamic evolution in LLM safety behavior, challenging the static alignment paradigm and extending it toward context-aware, adaptive safety modeling.

0 citationsRead paper

CAPTURE: A Benchmark and Evaluation for LVLMs in CAPTCHA Resolving

Dec 12, 2025

Existing CAPTCHA benchmarks lack LVLM-specific design, suffer from limited coverage of CAPTCHA types, and employ annotation schemes misaligned with LVLM capabilities. To address this, we introduce CAPTURE—the first dedicated evaluation benchmark for large vision-language models (LVLMs). CAPTURE comprises 25 fine-grained CAPTCHA subtypes across four major categories, sourced from 31 real-world service providers. It features a taxonomy-driven classification framework and an LVLM-adapted, fine-grained labeling scheme. Furthermore, we propose a multi-dimensional evaluation protocol explicitly designed to accommodate LVLM output characteristics (e.g., free-form text generation, multimodal reasoning). Extensive experiments reveal that state-of-the-art LVLMs achieve only 31.7% average accuracy on CAPTURE, exposing critical weaknesses in interference-robust text recognition and compositional reasoning. CAPTURE fills a fundamental gap in LVLM-specific security evaluation, providing a reproducible, systematic, and quantitative tool for model diagnostics and robustness research.

0 citationsRead paper

Oedipus and the Sphinx: Benchmarking and Improving Visual Language Models for Complex Graphic Reasoning

Aug 01, 2025

Current vision-language models (VLMs) lack systematic evaluation on complex graphical reasoning—encompassing spatial, relational, and abstract reasoning. To address this gap, we introduce ReasonBench, the first dedicated benchmark for graphical reasoning, comprising 1,613 authentic intelligence test items. We propose a dual-optimization framework: (1) DiaCoT, a hierarchical dialogue-based chain-of-thought method that enhances interpretability and stepwise reasoning; and (2) ReasonTune, a task-adaptive fine-tuning strategy that strengthens abstract structural modeling. Evaluated across 11 state-of-the-art VLMs, our approach achieves a 33.5% absolute accuracy improvement, revealing critical bottlenecks in structured graphical understanding. ReasonBench establishes a standardized, challenging evaluation platform and advances VLMs toward higher-order cognitive reasoning capabilities.

0 citationsRead paper

A Reward-driven Automated Webshell Malicious-code Generator for Red-teaming

May 30, 2025

Public WebShell malware datasets are scarce, and existing generation methods suffer from low payload diversity and high redundancy. Method: This paper proposes the first reward-driven, automated WebShell generation framework tailored for red-team operations. It introduces a novel LLM-based token standardization modeling system covering seven distinct obfuscation patterns and integrates supervised fine-tuning (SFT) with proximal policy optimization (PPO) for preference alignment—using malicious samples as “chosen” and benign code as “rejected.” Contribution/Results: The generated WebShells comprehensively cover all seven obfuscation patterns. Empirical evaluation shows a 32.7% average improvement in evasion success rate against mainstream detectors and a 61.4% reduction in redundancy, significantly outperforming baseline approaches.

0 citationsRead paper

Recent publications

Latest Papers

ShellForge: Adversarial Co-Evolution of Webshell Generation and Multi-View Detection for Robust Webshell Defense

Jan 28, 2026

This work addresses the challenge that existing webshell detection methods struggle to keep pace with rapidly evolving obfuscated variants and suffer from high false-positive rates on benign obfuscated scripts. To overcome these limitations, the authors propose an adversarial co-evolution framework in which a generator and a multi-perspective detector iteratively train each other to continuously enhance defensive capabilities. The approach innovatively leverages de-maliciousized samples as high-quality hard negatives and integrates supervised fine-tuning with preference-based reinforcement learning to generate highly evasive webshells. The detector combines multi-view features, including semantic compression of long strings, pruned abstract syntax tree (AST) structures, and Shannon entropy. Evaluated on the FWOID benchmark, the method achieves an F1-score of 0.981, and its generated samples attain a 0.939 evasion success rate against commercial VirusTotal engines, significantly reducing false positives while improving generalization.

0 citationsRead paper

MEEA: Mere Exposure Effect-Driven Confrontational Optimization for LLM Jailbreaking

Dec 21, 2025

Existing LLM safety alignment research predominantly relies on static boundary assumptions, neglecting the behavioral dynamics induced by contextual interactions—thereby limiting robustness and generalization. Method: We propose the first multi-round adversarial evaluation framework integrating psychology’s “mere-exposure effect,” employing semantically progressive, low-toxicity prompt chains to repeatedly expose black-box models and dynamically lower their safety thresholds. Our approach combines simulated annealing optimization, joint toxicity-similarity guidance, and strategic multi-turn interaction. Results: Experiments across major models—including GPT-4, Claude-3.5, and DeepSeek-R1—demonstrate an average attack success rate improvement exceeding 20% over seven baseline methods. This work provides the first empirical evidence of historical dependence and dynamic evolution in LLM safety behavior, challenging the static alignment paradigm and extending it toward context-aware, adaptive safety modeling.

0 citationsRead paper

CAPTURE: A Benchmark and Evaluation for LVLMs in CAPTCHA Resolving

Dec 12, 2025

Existing CAPTCHA benchmarks lack LVLM-specific design, suffer from limited coverage of CAPTCHA types, and employ annotation schemes misaligned with LVLM capabilities. To address this, we introduce CAPTURE—the first dedicated evaluation benchmark for large vision-language models (LVLMs). CAPTURE comprises 25 fine-grained CAPTCHA subtypes across four major categories, sourced from 31 real-world service providers. It features a taxonomy-driven classification framework and an LVLM-adapted, fine-grained labeling scheme. Furthermore, we propose a multi-dimensional evaluation protocol explicitly designed to accommodate LVLM output characteristics (e.g., free-form text generation, multimodal reasoning). Extensive experiments reveal that state-of-the-art LVLMs achieve only 31.7% average accuracy on CAPTURE, exposing critical weaknesses in interference-robust text recognition and compositional reasoning. CAPTURE fills a fundamental gap in LVLM-specific security evaluation, providing a reproducible, systematic, and quantitative tool for model diagnostics and robustness research.

0 citationsRead paper

Oedipus and the Sphinx: Benchmarking and Improving Visual Language Models for Complex Graphic Reasoning

Aug 01, 2025

Current vision-language models (VLMs) lack systematic evaluation on complex graphical reasoning—encompassing spatial, relational, and abstract reasoning. To address this gap, we introduce ReasonBench, the first dedicated benchmark for graphical reasoning, comprising 1,613 authentic intelligence test items. We propose a dual-optimization framework: (1) DiaCoT, a hierarchical dialogue-based chain-of-thought method that enhances interpretability and stepwise reasoning; and (2) ReasonTune, a task-adaptive fine-tuning strategy that strengthens abstract structural modeling. Evaluated across 11 state-of-the-art VLMs, our approach achieves a 33.5% absolute accuracy improvement, revealing critical bottlenecks in structured graphical understanding. ReasonBench establishes a standardized, challenging evaluation platform and advances VLMs toward higher-order cognitive reasoning capabilities.

0 citationsRead paper

A Reward-driven Automated Webshell Malicious-code Generator for Red-teaming

May 30, 2025

Public WebShell malware datasets are scarce, and existing generation methods suffer from low payload diversity and high redundancy. Method: This paper proposes the first reward-driven, automated WebShell generation framework tailored for red-team operations. It introduces a novel LLM-based token standardization modeling system covering seven distinct obfuscation patterns and integrates supervised fine-tuning (SFT) with proximal policy optimization (PPO) for preference alignment—using malicious samples as “chosen” and benign code as “rejected.” Contribution/Results: The generated WebShells comprehensively cover all seven obfuscation patterns. Empirical evaluation shows a 32.7% average improvement in evasion success rate against mainstream detectors and a 61.4% reduction in redundancy, significantly outperforming baseline approaches.

0 citationsRead paper