Selected work

Representative Papers

VectorizationLLM: Smart Vectorization Based AI Assistant

Jul 08, 2026

This work addresses the challenges engineering students face in grasping complex concepts such as intelligent vectorization, time–frequency analysis, piecewise functions, Fourier analysis, and differential equations in computational coursework. To support learning in CTEC 247, we developed an AI teaching assistant based on Google’s open-source large language model, uniquely integrating retrieval-augmented generation (RAG) with a pedagogical scaffolding mechanism. Through carefully engineered system prompts, the assistant delivers multimodal responses—comprising code, text, and visualizations—that provide conceptual explanations grounded in lecture notes rather than direct solutions. Deployed at New York Institute of Technology, the system significantly enhanced students’ comprehension of advanced computational topics and received positive instructional feedback, demonstrating its effectiveness and innovation in engineering education.

0 citationsRead paper

Zero Day Attacks: Novel Behaviour or Novel Vulnerability?

May 04, 2026

This study addresses a critical limitation in current behavior-based intrusion detection systems, which often fail to effectively identify zero-day attacks by misattributing their essence to behavioral novelty. Through a systematic analysis of real-world zero-day incidents over the past two decades, the authors construct a vulnerability taxonomy and empirically demonstrate that most zero-day exploits stem not from unprecedented behaviors but from undisclosed instances of known vulnerability patterns—particularly memory corruption vulnerabilities. The work critically reveals a fundamental misalignment between prevailing machine learning–based detection approaches and the actual mechanisms of contemporary attacks. It argues that a vulnerability-centric detection paradigm better aligns with practical defense requirements, especially given the recent surge in exploits specifically targeting defensive mechanisms themselves.

0 citationsRead paper

Evaluating Retrieval-Augmented Generation for Explainable Malware Analysis

May 04, 2026

This study investigates whether retrieval-augmented generation (RAG) genuinely enhances the performance of large language models (LLMs) in interpretable malware analysis, particularly when structured security data—such as VirusTotal reports—is already available. Through systematic comparisons of multiple LLMs with and without RAG, complemented by both qualitative and quantitative evaluations, the research reveals that when high-quality structured evidence is present, RAG often introduces distracting or weakly relevant context, leading to narrative noise and overly generalized explanations that degrade both accuracy and clarity. These findings challenge the prevailing assumption that RAG is universally beneficial and instead suggest that malware interpretation is fundamentally a signal extraction task rather than a knowledge retrieval problem, urging caution in deploying RAG for security-critical applications.

0 citationsRead paper

StableTTA: Training-Free Test-Time Adaptation that Improves Model Accuracy on ImageNet1K to 96%

Apr 06, 2026

This work addresses the instability and high computational and memory overhead commonly caused by conflicting aggregation strategies in ensemble methods. For the first time, it identifies this conflict as a critical issue and introduces a training-free, test-time adaptive aggregation mechanism that dynamically optimizes the fusion of multiple model outputs without requiring additional training or parameter updates. The proposed approach substantially enhances prediction stability and efficiency, achieving over 96% Top-1 accuracy on ImageNet-1K. Notably, it enables lightweight models to outperform Vision Transformers (ViTs) by 11.75% in accuracy while reducing model parameters by more than 95% and cutting computational costs by 89.1%.

0 citationsRead paper

SPARK: Secure Predictive Autoscaling for Robust Kubernetes

Mar 27, 2026

This work addresses the high response latency of traditional Kubernetes autoscaling mechanisms and their inability to distinguish between legitimate traffic bursts and DDoS attacks, which compromises both system availability and security. The authors propose a novel approach that integrates predictive autoscaling with eBPF-based kernel-level security policies to enable traffic-aware scaling decisions and Layer 7 isolation directly at the network layer. This is the first framework to combine predictive machine learning models with eBPF-driven network policies, extending Kubernetes’ Horizontal Pod Autoscaler (HPA) for coordinated optimization. Experimental results demonstrate that, compared to conventional reactive autoscaling, the proposed method reduces timeout errors by 32% under bursty traffic while ensuring immediate network convergence and secure isolation for newly scheduled pods.

0 citationsRead paper

Recent publications

Latest Papers

VectorizationLLM: Smart Vectorization Based AI Assistant

Jul 08, 2026

This work addresses the challenges engineering students face in grasping complex concepts such as intelligent vectorization, time–frequency analysis, piecewise functions, Fourier analysis, and differential equations in computational coursework. To support learning in CTEC 247, we developed an AI teaching assistant based on Google’s open-source large language model, uniquely integrating retrieval-augmented generation (RAG) with a pedagogical scaffolding mechanism. Through carefully engineered system prompts, the assistant delivers multimodal responses—comprising code, text, and visualizations—that provide conceptual explanations grounded in lecture notes rather than direct solutions. Deployed at New York Institute of Technology, the system significantly enhanced students’ comprehension of advanced computational topics and received positive instructional feedback, demonstrating its effectiveness and innovation in engineering education.

0 citationsRead paper

Zero Day Attacks: Novel Behaviour or Novel Vulnerability?

May 04, 2026

This study addresses a critical limitation in current behavior-based intrusion detection systems, which often fail to effectively identify zero-day attacks by misattributing their essence to behavioral novelty. Through a systematic analysis of real-world zero-day incidents over the past two decades, the authors construct a vulnerability taxonomy and empirically demonstrate that most zero-day exploits stem not from unprecedented behaviors but from undisclosed instances of known vulnerability patterns—particularly memory corruption vulnerabilities. The work critically reveals a fundamental misalignment between prevailing machine learning–based detection approaches and the actual mechanisms of contemporary attacks. It argues that a vulnerability-centric detection paradigm better aligns with practical defense requirements, especially given the recent surge in exploits specifically targeting defensive mechanisms themselves.

0 citationsRead paper

Evaluating Retrieval-Augmented Generation for Explainable Malware Analysis

May 04, 2026

This study investigates whether retrieval-augmented generation (RAG) genuinely enhances the performance of large language models (LLMs) in interpretable malware analysis, particularly when structured security data—such as VirusTotal reports—is already available. Through systematic comparisons of multiple LLMs with and without RAG, complemented by both qualitative and quantitative evaluations, the research reveals that when high-quality structured evidence is present, RAG often introduces distracting or weakly relevant context, leading to narrative noise and overly generalized explanations that degrade both accuracy and clarity. These findings challenge the prevailing assumption that RAG is universally beneficial and instead suggest that malware interpretation is fundamentally a signal extraction task rather than a knowledge retrieval problem, urging caution in deploying RAG for security-critical applications.

0 citationsRead paper

StableTTA: Training-Free Test-Time Adaptation that Improves Model Accuracy on ImageNet1K to 96%

Apr 06, 2026

This work addresses the instability and high computational and memory overhead commonly caused by conflicting aggregation strategies in ensemble methods. For the first time, it identifies this conflict as a critical issue and introduces a training-free, test-time adaptive aggregation mechanism that dynamically optimizes the fusion of multiple model outputs without requiring additional training or parameter updates. The proposed approach substantially enhances prediction stability and efficiency, achieving over 96% Top-1 accuracy on ImageNet-1K. Notably, it enables lightweight models to outperform Vision Transformers (ViTs) by 11.75% in accuracy while reducing model parameters by more than 95% and cutting computational costs by 89.1%.

0 citationsRead paper

SPARK: Secure Predictive Autoscaling for Robust Kubernetes

Mar 27, 2026

This work addresses the high response latency of traditional Kubernetes autoscaling mechanisms and their inability to distinguish between legitimate traffic bursts and DDoS attacks, which compromises both system availability and security. The authors propose a novel approach that integrates predictive autoscaling with eBPF-based kernel-level security policies to enable traffic-aware scaling decisions and Layer 7 isolation directly at the network layer. This is the first framework to combine predictive machine learning models with eBPF-driven network policies, extending Kubernetes’ Horizontal Pod Autoscaler (HPA) for coordinated optimization. Experimental results demonstrate that, compared to conventional reactive autoscaling, the proposed method reduces timeout errors by 32% under bursty traffic while ensuring immediate network convergence and secure isolation for newly scheduled pods.

0 citationsRead paper