Selected work

Representative Papers

Res-MIA: A Training-Free Resolution-Based Membership Inference Attack on Federated Learning Models

Jan 24, 2026

This work addresses the privacy risk in federated learning where black-box models may still leak membership information about training data. The authors propose a training-free black-box membership inference attack that exploits differences in model sensitivity to high-frequency details induced by varying input resolutions. By applying controlled downsampling and upsampling operations to simulate resolution degradation, the method analyzes the resulting divergence in prediction confidence decay between member and non-member samples to infer membership status. Notably, the attack requires neither shadow models, auxiliary datasets, nor additional training—only a small number of forward queries. Evaluated on a federated ResNet-18 trained on CIFAR-10, the approach achieves an AUC of 0.88, substantially outperforming existing training-free baselines while incurring minimal computational overhead.

0 citationsRead paper

A Non-Monotonic Relationship: An Empirical Analysis of Hybrid Quantum Classifiers for Unseen Ransomware Detection

Sep 09, 2025

Traditional machine learning methods suffer from poor generalization in detecting unknown ransomware, while quantum machine learning (QML) faces a dimensionality mismatch between classical data and current quantum hardware. Method: This paper proposes a PCA-VQC hybrid framework that applies principal component analysis (PCA) for dimensionality reduction prior to feeding features into a variational quantum classifier (VQC). Contribution/Results: Empirical evaluation demonstrates that the framework achieves 97.7% recall using only 12 qubits—significantly outperforming classical baselines. Notably, performance anomalously degrades at 4–8 qubits, revealing a non-monotonic coupling between training difficulty and information bottlenecks in QML. This work presents the first systematic validation of dimensionality reduction as an effective and scalable enabler of QML for real-world cybersecurity tasks. Moreover, it provides critical insights into the qubit-count–performance trade-off, informing principled design of quantum models for practical security applications.

0 citationsRead paper

Distillation of CNN Ensemble Results for Enhanced Long-Term Prediction of the ENSO Phenomenon

Sep 07, 2025

Current ENSO long-range forecasting commonly employs equal-weight ensemble averaging, overlooking inter-member skill heterogeneity and thus limiting prediction accuracy. To address this, we propose a posteriori assessment–based ensemble distillation method: leveraging outputs from a CNN-based ensemble forecasting system, we dynamically select high-skill ensemble subsets using the Niño3.4 index as reference, guided by RMSE and Pearson correlation coefficient. This approach departs from conventional arithmetic averaging. At an unprecedented 23-month lead time, it achieves substantial improvements—correlation increases to 0.71 (a relative gain of 172%, +0.43), and RMSE decreases by 22.5%. Notably, performance gains are most pronounced during climatological regime transitions, underscoring the critical importance of explicitly modeling ensemble skill heterogeneity for subseasonal-to-seasonal prediction.

0 citationsRead paper

Raising the Bar: An Asymptotic Comparison of Classical and Quantum Shortest Path Algorithms

Aug 16, 2025

This work reevaluates the asymptotic quantum advantage for the single-source shortest paths (SSSP) problem in light of recent advances in classical algorithms—specifically, the improved classical algorithm by Duan et al. Method: We conduct a systematic theoretical complexity comparison among Dijkstra’s algorithm, Duan’s classical algorithm, and quantum algorithms by Wesolowski–Piddock et al., across varying graph densities and shortest-path lengths $L$. We introduce a dynamic framework for assessing quantum advantage conditioned on structural parameters. Results: We establish that quantum algorithms achieve strict asymptotic advantage *only* when the shortest-path length satisfies $L = o(n^{1/3})$; for larger $L$, state-of-the-art classical algorithms consistently dominate. This demonstrates that progress in classical algorithm design is concretely narrowing the regime of quantum supremacy for SSSP. Moreover, our analysis provides a new paradigm for designing practical quantum algorithms tailored to real-world constraints—including path-scale limitations and graph topology—thereby shifting focus from unconditional speedups to context-aware quantum advantage.

0 citationsRead paper

Leveraging GNN to Enhance MEF Method in Predicting ENSO

Aug 10, 2025

ENSO long-term prediction accuracy is constrained by substantial performance heterogeneity among ensemble members, the absence of optimal weighting schemes, and noise contamination. To address these challenges, we propose a graph neural network (GNN)-driven ensemble optimization framework. First, we construct a similarity graph over ensemble members based on RMSE and temporal correlation; then, leveraging community detection, we identify high-performing subensembles (e.g., 20/80 splits). Finally, a model-agnostic GNN learns adaptive, member-specific weights for robust fusion. This approach uncovers emergent collective behavioral patterns among ensemble members, significantly enhancing forecast consistency and stability. Evaluated on multi-scale, long-lead ENSO prediction tasks, our method achieves an average 12.3% improvement in skill score over the baseline MEF ensemble, demonstrates superior noise suppression, and exhibits strong generalizability. The core innovation lies in integrating graph-structured representation learning with community-aware selection into multimodal ensemble optimization—establishing a novel paradigm for intelligent ensemble design in climate forecasting.

0 citationsRead paper

Recent publications

Latest Papers

Res-MIA: A Training-Free Resolution-Based Membership Inference Attack on Federated Learning Models

Jan 24, 2026

This work addresses the privacy risk in federated learning where black-box models may still leak membership information about training data. The authors propose a training-free black-box membership inference attack that exploits differences in model sensitivity to high-frequency details induced by varying input resolutions. By applying controlled downsampling and upsampling operations to simulate resolution degradation, the method analyzes the resulting divergence in prediction confidence decay between member and non-member samples to infer membership status. Notably, the attack requires neither shadow models, auxiliary datasets, nor additional training—only a small number of forward queries. Evaluated on a federated ResNet-18 trained on CIFAR-10, the approach achieves an AUC of 0.88, substantially outperforming existing training-free baselines while incurring minimal computational overhead.

0 citationsRead paper

A Non-Monotonic Relationship: An Empirical Analysis of Hybrid Quantum Classifiers for Unseen Ransomware Detection

Sep 09, 2025

Traditional machine learning methods suffer from poor generalization in detecting unknown ransomware, while quantum machine learning (QML) faces a dimensionality mismatch between classical data and current quantum hardware. Method: This paper proposes a PCA-VQC hybrid framework that applies principal component analysis (PCA) for dimensionality reduction prior to feeding features into a variational quantum classifier (VQC). Contribution/Results: Empirical evaluation demonstrates that the framework achieves 97.7% recall using only 12 qubits—significantly outperforming classical baselines. Notably, performance anomalously degrades at 4–8 qubits, revealing a non-monotonic coupling between training difficulty and information bottlenecks in QML. This work presents the first systematic validation of dimensionality reduction as an effective and scalable enabler of QML for real-world cybersecurity tasks. Moreover, it provides critical insights into the qubit-count–performance trade-off, informing principled design of quantum models for practical security applications.

0 citationsRead paper

Distillation of CNN Ensemble Results for Enhanced Long-Term Prediction of the ENSO Phenomenon

Sep 07, 2025

Current ENSO long-range forecasting commonly employs equal-weight ensemble averaging, overlooking inter-member skill heterogeneity and thus limiting prediction accuracy. To address this, we propose a posteriori assessment–based ensemble distillation method: leveraging outputs from a CNN-based ensemble forecasting system, we dynamically select high-skill ensemble subsets using the Niño3.4 index as reference, guided by RMSE and Pearson correlation coefficient. This approach departs from conventional arithmetic averaging. At an unprecedented 23-month lead time, it achieves substantial improvements—correlation increases to 0.71 (a relative gain of 172%, +0.43), and RMSE decreases by 22.5%. Notably, performance gains are most pronounced during climatological regime transitions, underscoring the critical importance of explicitly modeling ensemble skill heterogeneity for subseasonal-to-seasonal prediction.

0 citationsRead paper

Raising the Bar: An Asymptotic Comparison of Classical and Quantum Shortest Path Algorithms

Aug 16, 2025

This work reevaluates the asymptotic quantum advantage for the single-source shortest paths (SSSP) problem in light of recent advances in classical algorithms—specifically, the improved classical algorithm by Duan et al. Method: We conduct a systematic theoretical complexity comparison among Dijkstra’s algorithm, Duan’s classical algorithm, and quantum algorithms by Wesolowski–Piddock et al., across varying graph densities and shortest-path lengths $L$. We introduce a dynamic framework for assessing quantum advantage conditioned on structural parameters. Results: We establish that quantum algorithms achieve strict asymptotic advantage *only* when the shortest-path length satisfies $L = o(n^{1/3})$; for larger $L$, state-of-the-art classical algorithms consistently dominate. This demonstrates that progress in classical algorithm design is concretely narrowing the regime of quantum supremacy for SSSP. Moreover, our analysis provides a new paradigm for designing practical quantum algorithms tailored to real-world constraints—including path-scale limitations and graph topology—thereby shifting focus from unconditional speedups to context-aware quantum advantage.

0 citationsRead paper

Leveraging GNN to Enhance MEF Method in Predicting ENSO

Aug 10, 2025

ENSO long-term prediction accuracy is constrained by substantial performance heterogeneity among ensemble members, the absence of optimal weighting schemes, and noise contamination. To address these challenges, we propose a graph neural network (GNN)-driven ensemble optimization framework. First, we construct a similarity graph over ensemble members based on RMSE and temporal correlation; then, leveraging community detection, we identify high-performing subensembles (e.g., 20/80 splits). Finally, a model-agnostic GNN learns adaptive, member-specific weights for robust fusion. This approach uncovers emergent collective behavioral patterns among ensemble members, significantly enhancing forecast consistency and stability. Evaluated on multi-scale, long-lead ENSO prediction tasks, our method achieves an average 12.3% improvement in skill score over the baseline MEF ensemble, demonstrates superior noise suppression, and exhibits strong generalizability. The core innovation lies in integrating graph-structured representation learning with community-aware selection into multimodal ensemble optimization—establishing a novel paradigm for intelligent ensemble design in climate forecasting.

0 citationsRead paper