LiteRAG: Cost-Efficient Graph-Based Retrieval-Augmented Generation
该研究针对图检索中的高成本和低效问题，提出LiteRAG方法，通过算法探索与推理链构建降低查询延迟和成本，提高生成效率。
0 citationsRead paper
Selected work
Representative Papers
Exploring the connection between coding habits and cognitive styles in malware developers
This study addresses a critical gap in cybersecurity research by systematically examining the relationship between malware developers’ coding behaviors and their cognitive styles, an aspect largely overlooked in prior work that predominantly focuses on attack techniques. For the first time, code metrics are employed as behavioral proxies, integrating static application security testing (SAST) with software engineering measures—such as cyclomatic complexity, use of abstraction mechanisms, and vulnerability distributions—to comparatively analyze leaked malware samples against benign open-source projects. The findings reveal that malicious code tends to be smaller in scale, lacks documentation, exhibits higher function complexity, employs fewer abstraction mechanisms, and contains vulnerability types typically avoided by legitimate developers. These patterns reflect distinct motivational drivers, risk tolerance, and development priorities among malware authors, underscoring a strategy prioritizing efficiency and stealth over maintainability, thereby offering a novel empirical foundation for profiling cybercriminal behavior.
0 citationsRead paper
Emergent Self-Attention from Astrocyte-Gated Associative Memory Dynamics
This work addresses the limitations of traditional associative memory models, which suffer from degraded retrieval performance under high memory load and interference and lack a dynamical systems explanation for self-attention mechanisms. The authors propose a novel Hopfield-type associative memory model incorporating astrocyte-regulated neuronal gain, whose dynamics are governed by an entropy-regularized replicator equation. This formulation naturally yields softmax-normalized pattern similarity allocation over the gain simplex. The resulting coupled system exhibits global convergence and, for the first time, reveals self-attention as an emergent routing behavior modulated by astrocytes from a dynamical systems perspective. Experimental results demonstrate that the proposed model significantly outperforms classical Hopfield networks and existing neuro-glial baselines under conditions of high memory load and strong interference, achieving markedly higher retrieval accuracy.
0 citationsRead paper
A Critical Review on the Effectiveness and Privacy Threats of Membership Inference Attacks
This work proposes the first membership inference attack (MIA) threat assessment framework tailored to real-world deployment conditions, addressing the uncertainty surrounding whether MIAs truly constitute a substantive privacy threat beyond their common use as a proxy metric. Through a systematic literature review and theoretical analysis, the study conducts a unified evaluation of representative MIA methods under practical constraints. The findings reveal that most MIAs pose only limited privacy risks in realistic settings, suggesting that the prevailing practice of treating MIA success as a universal privacy measure may significantly overestimate actual threats. Consequently, this overestimation can lead to unnecessary sacrifices in model utility. By challenging the assumed efficacy of MIAs as default privacy indicators, this research establishes a new paradigm for more grounded and context-aware privacy evaluations.
0 citationsRead paper
Revisiting the LiRA Membership Inference Attack Under Realistic Assumptions
This work demonstrates that existing membership inference attacks, such as LiRA, are overestimated under unrealistic assumptions. The authors propose a more practical evaluation protocol to systematically assess the impact of anti-overfitting (AOF), transfer learning (TL), shadow model threshold calibration, low membership priors (π ≤ 10%), and sample-level reproducibility on LiRA’s effectiveness. Experimental results show that AOF substantially degrades attack performance, while TL further reduces success rates and simultaneously improves model accuracy. Under calibrated shadow models and low priors, LiRA exhibits a significant drop in positive predictive value (PPV), and the set of vulnerable samples at low false positive rates demonstrates poor reproducibility. This study is the first to reveal the practical limitations of LiRA under realistic training conditions.
0 citationsRead paper
Recent publications
Latest Papers
LiteRAG: Cost-Efficient Graph-Based Retrieval-Augmented Generation
该研究针对图检索中的高成本和低效问题，提出LiteRAG方法，通过算法探索与推理链构建降低查询延迟和成本，提高生成效率。
0 citationsRead paper
Exploring the connection between coding habits and cognitive styles in malware developers
This study addresses a critical gap in cybersecurity research by systematically examining the relationship between malware developers’ coding behaviors and their cognitive styles, an aspect largely overlooked in prior work that predominantly focuses on attack techniques. For the first time, code metrics are employed as behavioral proxies, integrating static application security testing (SAST) with software engineering measures—such as cyclomatic complexity, use of abstraction mechanisms, and vulnerability distributions—to comparatively analyze leaked malware samples against benign open-source projects. The findings reveal that malicious code tends to be smaller in scale, lacks documentation, exhibits higher function complexity, employs fewer abstraction mechanisms, and contains vulnerability types typically avoided by legitimate developers. These patterns reflect distinct motivational drivers, risk tolerance, and development priorities among malware authors, underscoring a strategy prioritizing efficiency and stealth over maintainability, thereby offering a novel empirical foundation for profiling cybercriminal behavior.
0 citationsRead paper
Emergent Self-Attention from Astrocyte-Gated Associative Memory Dynamics
This work addresses the limitations of traditional associative memory models, which suffer from degraded retrieval performance under high memory load and interference and lack a dynamical systems explanation for self-attention mechanisms. The authors propose a novel Hopfield-type associative memory model incorporating astrocyte-regulated neuronal gain, whose dynamics are governed by an entropy-regularized replicator equation. This formulation naturally yields softmax-normalized pattern similarity allocation over the gain simplex. The resulting coupled system exhibits global convergence and, for the first time, reveals self-attention as an emergent routing behavior modulated by astrocytes from a dynamical systems perspective. Experimental results demonstrate that the proposed model significantly outperforms classical Hopfield networks and existing neuro-glial baselines under conditions of high memory load and strong interference, achieving markedly higher retrieval accuracy.
0 citationsRead paper
A Critical Review on the Effectiveness and Privacy Threats of Membership Inference Attacks
This work proposes the first membership inference attack (MIA) threat assessment framework tailored to real-world deployment conditions, addressing the uncertainty surrounding whether MIAs truly constitute a substantive privacy threat beyond their common use as a proxy metric. Through a systematic literature review and theoretical analysis, the study conducts a unified evaluation of representative MIA methods under practical constraints. The findings reveal that most MIAs pose only limited privacy risks in realistic settings, suggesting that the prevailing practice of treating MIA success as a universal privacy measure may significantly overestimate actual threats. Consequently, this overestimation can lead to unnecessary sacrifices in model utility. By challenging the assumed efficacy of MIAs as default privacy indicators, this research establishes a new paradigm for more grounded and context-aware privacy evaluations.
0 citationsRead paper
Revisiting the LiRA Membership Inference Attack Under Realistic Assumptions
This work demonstrates that existing membership inference attacks, such as LiRA, are overestimated under unrealistic assumptions. The authors propose a more practical evaluation protocol to systematically assess the impact of anti-overfitting (AOF), transfer learning (TL), shadow model threshold calibration, low membership priors (π ≤ 10%), and sample-level reproducibility on LiRA’s effectiveness. Experimental results show that AOF substantially degrades attack performance, while TL further reduces success rates and simultaneously improves model accuracy. Under calibrated shadow models and low priors, LiRA exhibits a significant drop in positive predictive value (PPV), and the set of vulnerable samples at low false positive rates demonstrates poor reproducibility. This study is the first to reveal the practical limitations of LiRA under realistic training conditions.
0 citationsRead paper