DLG-Net：基于物理约束与主动诱饵的离散液态图网络防幻觉架构
DLG-Net: A Physics-Constrained Discrete Liquid Graph Network with Active Canary Traps for Anti-Hallucination
摘要
针对极端边缘环境下AI系统易产生系统性误判与遭受对抗篡改的痛点，本文提出了一种物理约束的离散液态图（DLG）防幻觉安全底座。该架构构建了基于物理公理的确定性门控，从源头切断假阳性；首创了异构质询与主动诱饵反侦察机制，有效抵御数据投毒；并引入动态突触修剪与优雅降级策略。在理论与数学层面，本文结合时变拓扑与离散李雅普诺夫理论，严格证明了系统的有界收敛与渐近稳定性。在诸如现场质谱分析、痕量物质检测等安全攸关领域，边缘端 AI 系统极易遭受对抗性数据投毒与系统性幻觉的干扰，进而产生严重的假阳性误报。基于MassBank真实质谱数据集的压力测试表明，该架构在各类对抗攻击下F1-Score均超过0.99，将误报率严格控制在1.3%，且在纯CPU环境下推理延迟仅5.31毫秒。该研究为下一代可信边缘人工智能提供了一种具备绝对安全边界的新范式。
Abstract
To address systematic misclassification and adversarial tampering in extreme edge environments, this paper proposes a Physics-Constrained Discrete Liquid Graph (DLG) anti-hallucination security base. The architecture employs a deterministic physical gating to eliminate false positives, introduces heterogeneous cross-examination with Active Canary Traps to resist data poisoning, and integrates dynamic synaptic pruning with graceful degradation. Theoretically, by combining time-varying topology with discrete Lyapunov analysis, we strictly prove bounded convergence and asymptotic stability. In safety-critical domains such as on-site mass spectrometry (MS) and real-time trace chemical detection, edge AI systems are highly vulnerable to adversarial data poisoning and systematic hallucinations, leading to severe false-positive alarms. Evaluations on the MassBank dataset demonstrate that the DLG architecture maintains an F1-Score above 0.99 under various adversarial attacks, restricts the false positive rate to 1.3%, and achieves a pure CPU inference latency of merely 5.31 ms. This work provides a novel paradigm for next-generation trustworthy edge AI.关键词
边缘计算/AI安全/离散液态图神经网络/物理约束/拜占庭容错/李雅普诺夫稳定性/防幻觉架构/质谱分析Key words
Edge Computing/AI Safety/Discrete Liquid Graph Neural Network/Physics Constraints/Byzantine Fault Tolerance/Lyapunov Stability/Anti-Hallucination Architecture/Mass Spectrometry引用本文复制引用
李欣怡,程平.DLG-Net：基于物理约束与主动诱饵的离散液态图网络防幻觉架构[EB/OL].(2026-11-08)[2026-11-11].https://chinaxiv.org/abs/202611.00064.学科分类
计算技术、计算机技术
