Selected work

Representative Papers

Machine Unlearning for the XGBoost Model with Network Intrusion Datasets

Jun 17, 2026

This work addresses the gap in machine unlearning research by introducing XGBoost-Forget, the first unlearning method tailored for XGBoost models in tabular network intrusion detection scenarios. Unlike existing approaches primarily designed for deep learning and image data, XGBoost-Forget efficiently removes specified intrusion data points without requiring full model retraining. The method incorporates a customized unlearning mechanism specifically designed for tabular network traffic data and is validated on real-world datasets such as IoT-23 and GeNIS. Experimental results demonstrate that XGBoost-Forget achieves substantial gains in unlearning efficiency while preserving predictive performance nearly equivalent to that of the original model, thereby offering a practical unlearning solution for security applications driven by tabular data.

0 citationsRead paper

Influence Factors on RAG Poisoning

Jun 09, 2026

This study systematically investigates the sources of vulnerability in Retrieval-Augmented Generation (RAG) systems under poisoning attacks. Through a comprehensive full-factorial experiment encompassing 432 configurations, it evaluates the impact of datasets, retriever types (dense, graph-based, and BM25), retrieval depth, knowledge base composition, chunking strategies, and generation models on system robustness. The findings reveal that RAG’s susceptibility arises from complex interactions among retrieval, generation, and knowledge base configurations rather than from any single component’s deficiency. Dense and graph-based retrievers significantly outperform BM25, while increasing retrieval depth or replicating poisoned content across multiple knowledge sources substantially elevates attack success rates. Conversely, incorporating clean data from diverse sources effectively mitigates such attacks. This work is the first to uncover the key factors governing RAG’s robustness against poisoning and their underlying coupling mechanisms.

0 citationsRead paper

Towards Privacy-Preserving Federated Learning using Hybrid Homomorphic Encryption

Mar 27, 2026

This work addresses a critical vulnerability in existing homomorphic encryption–based federated learning (HHE-FL) systems, which rely on a single key pair and are thus susceptible to privacy breaches by malicious clients. To mitigate this risk, the paper introduces two novel key protection mechanisms—homomorphic masking and RSA-based public-key encapsulation—thereby elevating the security of HHE-FL to a strong threat model resilient against adversarial participants for the first time. The proposed methods are implemented within the Flower framework using the PASTA/BFV hybrid homomorphic encryption scheme. Experimental evaluation on MNIST with 12 clients demonstrates that both mechanisms preserve model accuracy while incurring minimal overhead: homomorphic masking adds negligible computational cost, and RSA encapsulation introduces only modest communication and runtime overhead.

0 citationsRead paper

Machine Learning Transferability for Malware Detection

Mar 27, 2026

This work addresses the limited cross-dataset generalization of existing malware detection methods, which stems from inconsistent feature representations in public datasets and susceptibility to distributional shifts. To mitigate this, the authors propose a preprocessing framework that unifies the feature space based on EMBERv2, integrates the BODMAS dataset, and introduces an ERMDS regularization-enhanced training strategy to enable joint training on multi-source PE files within the EMBER framework. Experimental results demonstrate that the proposed approach significantly improves transfer detection performance across multiple heterogeneous external test sets—including TRITIUM, INFERNO, and SOREL-20M—exhibiting superior generalization capability and robustness compared to baseline methods.

0 citationsRead paper

Axle Sensor Fusion for Online Continual Wheel Fault Detection in Wayside Railway Monitoring

Feb 17, 2026

This work addresses the challenges of online wheelset fault detection in railways, where model performance degrades under varying operating conditions, reliance on handcrafted features persists, and labeled data are scarce. To this end, the authors propose a semantic-aware, label-efficient continual learning framework that integrates semantic metadata—such as axle count, wheel position, and strain—derived from accelerometer and fiber Bragg grating sensors. The approach combines unsupervised variational autoencoder representations with a lightweight gradient-boosting classifier and incorporates a replay-based continual learning mechanism. This is the first method to jointly optimize semantic metadata and unsupervised deep representations, enabling effective adaptation to unseen operational conditions with minimal labeled data while mitigating catastrophic forgetting. Experiments demonstrate accurate detection of subtle defects like spalls and polygonization, maintaining high stability across variations in train type, speed, load, and track conditions, using only a single accelerometer and strain gauge for deployment.

0 citationsRead paper

Recent publications

Latest Papers

Machine Unlearning for the XGBoost Model with Network Intrusion Datasets

Jun 17, 2026

This work addresses the gap in machine unlearning research by introducing XGBoost-Forget, the first unlearning method tailored for XGBoost models in tabular network intrusion detection scenarios. Unlike existing approaches primarily designed for deep learning and image data, XGBoost-Forget efficiently removes specified intrusion data points without requiring full model retraining. The method incorporates a customized unlearning mechanism specifically designed for tabular network traffic data and is validated on real-world datasets such as IoT-23 and GeNIS. Experimental results demonstrate that XGBoost-Forget achieves substantial gains in unlearning efficiency while preserving predictive performance nearly equivalent to that of the original model, thereby offering a practical unlearning solution for security applications driven by tabular data.

0 citationsRead paper

Influence Factors on RAG Poisoning

Jun 09, 2026

This study systematically investigates the sources of vulnerability in Retrieval-Augmented Generation (RAG) systems under poisoning attacks. Through a comprehensive full-factorial experiment encompassing 432 configurations, it evaluates the impact of datasets, retriever types (dense, graph-based, and BM25), retrieval depth, knowledge base composition, chunking strategies, and generation models on system robustness. The findings reveal that RAG’s susceptibility arises from complex interactions among retrieval, generation, and knowledge base configurations rather than from any single component’s deficiency. Dense and graph-based retrievers significantly outperform BM25, while increasing retrieval depth or replicating poisoned content across multiple knowledge sources substantially elevates attack success rates. Conversely, incorporating clean data from diverse sources effectively mitigates such attacks. This work is the first to uncover the key factors governing RAG’s robustness against poisoning and their underlying coupling mechanisms.

0 citationsRead paper

Towards Privacy-Preserving Federated Learning using Hybrid Homomorphic Encryption

Mar 27, 2026

This work addresses a critical vulnerability in existing homomorphic encryption–based federated learning (HHE-FL) systems, which rely on a single key pair and are thus susceptible to privacy breaches by malicious clients. To mitigate this risk, the paper introduces two novel key protection mechanisms—homomorphic masking and RSA-based public-key encapsulation—thereby elevating the security of HHE-FL to a strong threat model resilient against adversarial participants for the first time. The proposed methods are implemented within the Flower framework using the PASTA/BFV hybrid homomorphic encryption scheme. Experimental evaluation on MNIST with 12 clients demonstrates that both mechanisms preserve model accuracy while incurring minimal overhead: homomorphic masking adds negligible computational cost, and RSA encapsulation introduces only modest communication and runtime overhead.

0 citationsRead paper

Machine Learning Transferability for Malware Detection

Mar 27, 2026

This work addresses the limited cross-dataset generalization of existing malware detection methods, which stems from inconsistent feature representations in public datasets and susceptibility to distributional shifts. To mitigate this, the authors propose a preprocessing framework that unifies the feature space based on EMBERv2, integrates the BODMAS dataset, and introduces an ERMDS regularization-enhanced training strategy to enable joint training on multi-source PE files within the EMBER framework. Experimental results demonstrate that the proposed approach significantly improves transfer detection performance across multiple heterogeneous external test sets—including TRITIUM, INFERNO, and SOREL-20M—exhibiting superior generalization capability and robustness compared to baseline methods.

0 citationsRead paper

Axle Sensor Fusion for Online Continual Wheel Fault Detection in Wayside Railway Monitoring

Feb 17, 2026

This work addresses the challenges of online wheelset fault detection in railways, where model performance degrades under varying operating conditions, reliance on handcrafted features persists, and labeled data are scarce. To this end, the authors propose a semantic-aware, label-efficient continual learning framework that integrates semantic metadata—such as axle count, wheel position, and strain—derived from accelerometer and fiber Bragg grating sensors. The approach combines unsupervised variational autoencoder representations with a lightweight gradient-boosting classifier and incorporates a replay-based continual learning mechanism. This is the first method to jointly optimize semantic metadata and unsupervised deep representations, enabling effective adaptation to unseen operational conditions with minimal labeled data while mitigating catastrophic forgetting. Experiments demonstrate accurate detection of subtle defects like spalls and polygonization, maintaining high stability across variations in train type, speed, load, and track conditions, using only a single accelerometer and strain gauge for deployment.

0 citationsRead paper