Selected work

Representative Papers

MIRANDA: MId-feature RANk-adversarial Domain Adaptation toward climate change-robust ecological forecasting with deep learning

Apr 01, 2026

This study addresses the challenge of temporal domain shift—encompassing both covariate and label shifts—induced by climate change, which severely degrades the generalization of deep learning models in plant phenology prediction. To mitigate this issue, the authors propose MIRANDA, a novel approach that introduces rank-based adversarial regularization at intermediate feature layers to learn year-invariant meteorological representations. Departing from conventional binary domain discrimination, MIRANDA employs a ranking objective that simultaneously handles continuous temporal domain evolution and label distribution shifts. Evaluated on a national-scale dataset spanning 70 years with 67,800 records, MIRANDA substantially enhances model robustness to climate change and significantly narrows the performance gap with process-based mechanistic models.

0 citationsRead paper

IHearYou: Linking Acoustic Features to DSM-5 Depressive Behavior Indicators

Nov 17, 2025

Depression diagnosis relies heavily on subjective self-reports, limiting objective behavioral assessment. To address this, we propose a privacy-preserving, interpretable, speech-driven automated detection framework. Our method employs passive in-home speech acquisition and establishes a structured mapping between acoustic features and DSM-5–defined depressive behavioral indicators. It adopts a localized real-time processing architecture, incorporating gender-stratified validation and false discovery rate (FDR) correction for rigorous statistical control and reproducibility. We conduct end-to-end evaluation on the DAIC-WOZ dataset and TESS audio streams. Experiments reveal consistent, cross-dataset feature–indicator associations and achieve millisecond-scale inference on commodity laptops. Our key contribution is the first explicit integration of clinical diagnostic criteria—specifically DSM-5 symptom domains—into the speech analysis pipeline, thereby bridging the interpretability gap between opaque machine learning models and clinical psychiatric practice.

0 citationsRead paper

Distributed Pulse-Wave Simulator for DDoS Dataset Generation

Nov 16, 2025

Distributed pulse-wave DDoS attacks exhibit transient behavior, spatiotemporal synchronization, and cross-domain distribution, enabling evasion of conventional detection methods and lacking publicly available multi-domain evolutionary datasets. Method: We present the first open-source distributed pulse-wave DDoS traffic simulator, implemented in ns-3 to enable coordinated modeling across multiple autonomous systems (ASes). It supports MPI-based parallel simulation and YAML-driven fine-grained attack configuration. Innovatively integrating multi-point synchronized packet capture and a custom scheduling module, it generates reproducible, multi-perspective datasets exhibiting natural fingerprint variation. Contribution/Results: This work fills a critical gap in distributed pulse-wave attack data. Empirical evaluation confirms statistically significant inter-domain fingerprint divergence—even under identical aggregate traffic rates—enabling robust early detection, forensic attribution, and evaluation of distributed defense mechanisms. The simulator establishes a benchmark platform and provides a scalable, synchronized simulation framework for research and development.

0 citationsRead paper

FEST: A Unified Framework for Evaluating Synthetic Tabular Data

Aug 22, 2025

Existing synthetic tabular data evaluation methods lack a systematic framework that jointly addresses privacy preservation and data utility. This paper proposes FEST—the first unified, scalable synthetic data evaluation framework—designed to holistically quantify privacy–utility trade-offs across multiple dimensions. FEST uniquely integrates adversarial privacy metrics (e.g., membership inference, attribute inference) with distance-based privacy metrics (e.g., Jensen–Shannon divergence), while simultaneously assessing statistical fidelity (distributional similarity, correlation structure) and machine learning utility (downstream task performance). Built upon generative modeling principles, the framework is accompanied by an open-source Python library. Extensive validation on multiple benchmark datasets demonstrates its effectiveness and robustness. FEST enables standardized, cross-model comparative analysis of privacy–utility trade-offs, thereby providing a principled, reproducible tool for trustworthy synthetic data evaluation.

0 citationsRead paper

A Crowdsensing Intrusion Detection Dataset For Decentralized Federated Learning Models

Jul 17, 2025

This work addresses the privacy-efficiency trade-off in malware detection for Internet-of-Things (IoT) crowdsourced sensing. To this end, we propose a decentralized federated learning (DFL)-based framework that preserves data locality while enabling collaborative model training. As foundational support, we introduce the first DFL-oriented IoT intrusion detection dataset, comprising 342,106 samples derived from 30-second sliding windows and featuring multi-source behavioral attributes—including system calls, file operations, and resource utilization—across benign applications and eight major malware families. Extensive experiments under diverse network topologies and non-IID data distributions demonstrate that DFL consistently outperforms centralized federated learning (CFL) in both model accuracy and communication efficiency, without compromising data privacy. Our contribution establishes a new paradigm for lightweight, scalable, and privacy-enhancing IoT security analytics, accompanied by a publicly available benchmark dataset and empirical evaluation framework.

0 citationsRead paper

Recent publications

Latest Papers

MIRANDA: MId-feature RANk-adversarial Domain Adaptation toward climate change-robust ecological forecasting with deep learning

Apr 01, 2026

This study addresses the challenge of temporal domain shift—encompassing both covariate and label shifts—induced by climate change, which severely degrades the generalization of deep learning models in plant phenology prediction. To mitigate this issue, the authors propose MIRANDA, a novel approach that introduces rank-based adversarial regularization at intermediate feature layers to learn year-invariant meteorological representations. Departing from conventional binary domain discrimination, MIRANDA employs a ranking objective that simultaneously handles continuous temporal domain evolution and label distribution shifts. Evaluated on a national-scale dataset spanning 70 years with 67,800 records, MIRANDA substantially enhances model robustness to climate change and significantly narrows the performance gap with process-based mechanistic models.

0 citationsRead paper

IHearYou: Linking Acoustic Features to DSM-5 Depressive Behavior Indicators

Nov 17, 2025

Depression diagnosis relies heavily on subjective self-reports, limiting objective behavioral assessment. To address this, we propose a privacy-preserving, interpretable, speech-driven automated detection framework. Our method employs passive in-home speech acquisition and establishes a structured mapping between acoustic features and DSM-5–defined depressive behavioral indicators. It adopts a localized real-time processing architecture, incorporating gender-stratified validation and false discovery rate (FDR) correction for rigorous statistical control and reproducibility. We conduct end-to-end evaluation on the DAIC-WOZ dataset and TESS audio streams. Experiments reveal consistent, cross-dataset feature–indicator associations and achieve millisecond-scale inference on commodity laptops. Our key contribution is the first explicit integration of clinical diagnostic criteria—specifically DSM-5 symptom domains—into the speech analysis pipeline, thereby bridging the interpretability gap between opaque machine learning models and clinical psychiatric practice.

0 citationsRead paper

Distributed Pulse-Wave Simulator for DDoS Dataset Generation

Nov 16, 2025

Distributed pulse-wave DDoS attacks exhibit transient behavior, spatiotemporal synchronization, and cross-domain distribution, enabling evasion of conventional detection methods and lacking publicly available multi-domain evolutionary datasets. Method: We present the first open-source distributed pulse-wave DDoS traffic simulator, implemented in ns-3 to enable coordinated modeling across multiple autonomous systems (ASes). It supports MPI-based parallel simulation and YAML-driven fine-grained attack configuration. Innovatively integrating multi-point synchronized packet capture and a custom scheduling module, it generates reproducible, multi-perspective datasets exhibiting natural fingerprint variation. Contribution/Results: This work fills a critical gap in distributed pulse-wave attack data. Empirical evaluation confirms statistically significant inter-domain fingerprint divergence—even under identical aggregate traffic rates—enabling robust early detection, forensic attribution, and evaluation of distributed defense mechanisms. The simulator establishes a benchmark platform and provides a scalable, synchronized simulation framework for research and development.

0 citationsRead paper

FEST: A Unified Framework for Evaluating Synthetic Tabular Data

Aug 22, 2025

Existing synthetic tabular data evaluation methods lack a systematic framework that jointly addresses privacy preservation and data utility. This paper proposes FEST—the first unified, scalable synthetic data evaluation framework—designed to holistically quantify privacy–utility trade-offs across multiple dimensions. FEST uniquely integrates adversarial privacy metrics (e.g., membership inference, attribute inference) with distance-based privacy metrics (e.g., Jensen–Shannon divergence), while simultaneously assessing statistical fidelity (distributional similarity, correlation structure) and machine learning utility (downstream task performance). Built upon generative modeling principles, the framework is accompanied by an open-source Python library. Extensive validation on multiple benchmark datasets demonstrates its effectiveness and robustness. FEST enables standardized, cross-model comparative analysis of privacy–utility trade-offs, thereby providing a principled, reproducible tool for trustworthy synthetic data evaluation.

0 citationsRead paper

A Crowdsensing Intrusion Detection Dataset For Decentralized Federated Learning Models

Jul 17, 2025

This work addresses the privacy-efficiency trade-off in malware detection for Internet-of-Things (IoT) crowdsourced sensing. To this end, we propose a decentralized federated learning (DFL)-based framework that preserves data locality while enabling collaborative model training. As foundational support, we introduce the first DFL-oriented IoT intrusion detection dataset, comprising 342,106 samples derived from 30-second sliding windows and featuring multi-source behavioral attributes—including system calls, file operations, and resource utilization—across benign applications and eight major malware families. Extensive experiments under diverse network topologies and non-IID data distributions demonstrate that DFL consistently outperforms centralized federated learning (CFL) in both model accuracy and communication efficiency, without compromising data privacy. Our contribution establishes a new paradigm for lightweight, scalable, and privacy-enhancing IoT security analytics, accompanied by a publicly available benchmark dataset and empirical evaluation framework.

0 citationsRead paper