When Review Alone No Longer Scales: Layered Supervision in AI-Assisted Software Engineering
研究探讨了AI辅助软件工程中，通过分层监督（包括预防性、可执行性和人工监督）来应对传统代码审查等控制机制面临的压力。
0 citationsRead paper
Selected work
Representative Papers
Physics Informed Human Posture Estimation Based on 3D Landmarks from Monocular RGB-Videos
To address anatomical implausibility and insufficient robustness in monocular RGB video-based 3D human pose estimation, this paper proposes a real-time, anatomy-aware optimization framework integrating physics-based priors with deep learning. Methodologically, it refines BlazePose’s 2D/3D keypoint outputs by incorporating subject-specific bone-length modeling, biomechanical constraints—including joint angle limits and kinematic connectivity—and a bone-length physical penalty term. An adaptive-confidence Kalman filter dynamically calibrates anatomical parameters without retraining the base model. Evaluated on Physio2.2M, the method reduces 3D MPJPE by 10.2% and joint angle error by 16.6%, while enabling real-time inference on consumer-grade edge devices with on-device privacy preservation. The core contribution is a lightweight, anatomy-consistency-driven optimization paradigm that significantly enhances reliability and practicality for clinical applications—such as physical therapy and sports coaching—without compromising computational efficiency.
0 citationsRead paper
Multi-Agent Reinforcement Learning in Cybersecurity: From Fundamentals to Applications
To address the limitations of existing network defense approaches in scalability and adversarial robustness, this paper proposes a decentralized dynamic defense framework based on multi-agent reinforcement learning (MARL), targeting two core challenges: intrusion detection and lateral movement containment. Methodologically, it introduces the first systematic integration of Autonomous Intelligent Cyber Defense Agents (AICA) with the Cyber Gym cybersecurity simulation platform to establish a distributed collaborative decision-making architecture, augmented by adversarial training to enhance robustness. The contributions are threefold: (1) establishing a principled MARL-driven technical roadmap for cyber defense; (2) empirically validating the framework’s effectiveness and generalizability in dynamic threat response, cooperative containment, and cross-domain coordination; and (3) providing a novel paradigm for scalable, adaptive, and interference-resilient automated defense in realistic operational environments.
0 citationsRead paper
Securing RAG: A Risk Assessment and Mitigation Framework
Integrating sensitive data into Retrieval-Augmented Generation (RAG) systems introduces novel security and privacy risks across the end-to-end pipeline—spanning data preprocessing, storage, retrieval, and large language model (LLM) generation. Method: We systematically identify and characterize the full RAG attack surface, proposing the first RAG-specific attack surface definition methodology. Our structured governance framework integrates domain-specific RAG properties with established standards—including ISO/IEC 27001 and NIST SP 800-53—enabling bidirectional mapping between security controls and compliance requirements. We conduct rigorous risk modeling, attack tree analysis, and RAG pipeline security auditing. Contribution/Results: The work yields a practical, actionable security checklist, a mitigation strategy matrix, and an implementation guide. It directly supports enterprise RAG deployments in achieving compliance with China’s Multi-Level Protection Scheme (MLPS) Level 3 and the GDPR—thereby bridging critical theoretical and practical gaps in RAG security governance.
0 citationsRead paper
Recent publications
Latest Papers
When Review Alone No Longer Scales: Layered Supervision in AI-Assisted Software Engineering
研究探讨了AI辅助软件工程中，通过分层监督（包括预防性、可执行性和人工监督）来应对传统代码审查等控制机制面临的压力。
0 citationsRead paper
Physics Informed Human Posture Estimation Based on 3D Landmarks from Monocular RGB-Videos
To address anatomical implausibility and insufficient robustness in monocular RGB video-based 3D human pose estimation, this paper proposes a real-time, anatomy-aware optimization framework integrating physics-based priors with deep learning. Methodologically, it refines BlazePose’s 2D/3D keypoint outputs by incorporating subject-specific bone-length modeling, biomechanical constraints—including joint angle limits and kinematic connectivity—and a bone-length physical penalty term. An adaptive-confidence Kalman filter dynamically calibrates anatomical parameters without retraining the base model. Evaluated on Physio2.2M, the method reduces 3D MPJPE by 10.2% and joint angle error by 16.6%, while enabling real-time inference on consumer-grade edge devices with on-device privacy preservation. The core contribution is a lightweight, anatomy-consistency-driven optimization paradigm that significantly enhances reliability and practicality for clinical applications—such as physical therapy and sports coaching—without compromising computational efficiency.
0 citationsRead paper
Multi-Agent Reinforcement Learning in Cybersecurity: From Fundamentals to Applications
To address the limitations of existing network defense approaches in scalability and adversarial robustness, this paper proposes a decentralized dynamic defense framework based on multi-agent reinforcement learning (MARL), targeting two core challenges: intrusion detection and lateral movement containment. Methodologically, it introduces the first systematic integration of Autonomous Intelligent Cyber Defense Agents (AICA) with the Cyber Gym cybersecurity simulation platform to establish a distributed collaborative decision-making architecture, augmented by adversarial training to enhance robustness. The contributions are threefold: (1) establishing a principled MARL-driven technical roadmap for cyber defense; (2) empirically validating the framework’s effectiveness and generalizability in dynamic threat response, cooperative containment, and cross-domain coordination; and (3) providing a novel paradigm for scalable, adaptive, and interference-resilient automated defense in realistic operational environments.
0 citationsRead paper
Securing RAG: A Risk Assessment and Mitigation Framework
Integrating sensitive data into Retrieval-Augmented Generation (RAG) systems introduces novel security and privacy risks across the end-to-end pipeline—spanning data preprocessing, storage, retrieval, and large language model (LLM) generation. Method: We systematically identify and characterize the full RAG attack surface, proposing the first RAG-specific attack surface definition methodology. Our structured governance framework integrates domain-specific RAG properties with established standards—including ISO/IEC 27001 and NIST SP 800-53—enabling bidirectional mapping between security controls and compliance requirements. We conduct rigorous risk modeling, attack tree analysis, and RAG pipeline security auditing. Contribution/Results: The work yields a practical, actionable security checklist, a mitigation strategy matrix, and an implementation guide. It directly supports enterprise RAG deployments in achieving compliance with China’s Multi-Level Protection Scheme (MLPS) Level 3 and the GDPR—thereby bridging critical theoretical and practical gaps in RAG security governance.
0 citationsRead paper