Selected work

Representative Papers

Agentic AI-Powered Re-Identification: An Emerging, Scalable Threat to Mobility Microdata Privacy

Jun 26, 2026

This work addresses the severe re-identification privacy risks posed by fine-grained location data collected by commercial data brokers, which traditional attacks struggle to scale due to reliance on manual analysis. The paper proposes the first end-to-end automated re-identification framework, leveraging large language model agents to autonomously harvest publicly available online information and integrate public records, social media profiles, and spatiotemporal trajectory matching algorithms—enabling large-scale identity inference without human intervention. Evaluated on a simulated dataset containing home and workplace address anchors, the method successfully re-identifies 18 out of 43 individuals (41.9%), achieving 72% accuracy among identifiable subjects. This study demonstrates, for the first time in a realistic setting, the feasibility of fully automated, low-cost, and highly efficient re-identification from mobile microdata, fundamentally challenging the conventional paradigm that depends on expert involvement.

0 citationsRead paper

Log-Ratio Propagation on the Simplex: A Theory of Cellwise Contamination for Compositional Data

May 29, 2026

This study addresses the ill-posedness of conventional Euclidean robust methods on the simplex when a single compositional component is contaminated, which induces a global shift in log-ratio coordinates. Building upon a scale-invariant multiplicative perturbation model, the work establishes the first theoretical framework for cellwise contamination in compositional data, proving that contamination in one component manifests as a rank-one shift in log-ratio space. The authors introduce a contamination propagation theorem and an influence function–based diagnostic fingerprint. By leveraging centered log-ratio transformation, isometric log-ratio coordinates, and contrast matrix analysis, they quantify the cellwise breakdown points of MCD, S-, τ-, and coordinatewise M-estimators, revealing a reduction by a factor of $(D-1)/D$ compared to their Euclidean counterparts. Furthermore, they demonstrate that the variability matrix’s influence function precisely identifies contaminated components, thereby laying the theoretical foundation for cellwise robust analysis of compositional data.

0 citationsRead paper

Low-Magnification SEM May Suffice: Interpretable Deep Learning for Multi-Scale Fracture-Cause Classification in Zirconia-Toughened Alumina

May 28, 2026

This study addresses the limitations of conventional fracture analysis for ceramic implants, which relies on high-magnification scanning electron microscopy (SEM)—a time-consuming and subjective process. The authors propose an interpretable deep learning approach based on Vision Transformers to automatically classify fracture origins in zirconia-toughened alumina ceramics—specifically green-state, hard-machining, and material defects—using multi-scale SEM images. They demonstrate for the first time that low-magnification SEM (50×) contains sufficient diagnostic information, achieving classification performance comparable to that of high-magnification images. By integrating Grad-CAM, the model provides spatially interpretable predictions aligned with established fractographic standards. Despite severe class imbalance, the method attains 0.907 accuracy and 0.888 macro F1-score, enabling effective low-magnification prescreening and substantially reducing reliance on high-magnification SEM.

0 citationsRead paper

Synthetic Image Detection with CLIP: Understanding and Assessing Predictive Cues

Feb 12, 2026

Recent generative models produce near-photorealistic images, challenging the trustworthiness of photographs. Synthetic image detection (SID) has thus become an important area of research. Prior work has highlighted how synthetic images differ from real photographs--unfortunately, SID methods often struggle to generalize to novel generative models and often perform poorly in practical settings. CLIP, a foundational vision-language model which yields semantically rich image-text embeddings, shows strong accuracy and generalization for SID. Yet, the underlying relevant cues embedded in CLIP-features remain unknown. It is unclear, whether CLIP-based detectors simply detect strong visual artifacts or exploit subtle semantic biases, both of which would render them useless in practical settings or on generative models of high quality. We introduce SynthCLIC, a paired dataset of real photographs and high-quality synthetic counterparts from recent diffusion models, designed to reduce semantic bias in SID. Using an interpretable linear head with de-correlated activations and a text-grounded concept-model, we analyze what CLIP-based detectors learn. CLIP-based linear detectors reach 0.96 mAP on a GAN-based benchmark but only 0.92 on our high-quality diffusion dataset SynthCLIC, and generalization across generator families drops to as low as 0.37 mAP. We find that the detectors primarily rely on high-level photographic attributes (e.g., minimalist style, lens flare, or depth layering), rather than overt generator-specific artifacts. CLIP-based detectors perform well overall but generalize unevenly across diverse generative architectures. This highlights the need for continual model updates and broader training exposure, while reinforcing CLIP-based approaches as a strong foundation for more universal, robust SID.

0 citationsRead paper

RAPID: Risk of Attribute Prediction-Induced Disclosure in Synthetic Microdata

Feb 09, 2026

This work proposes RAPID, a novel attribute inference risk metric designed to better reflect realistic attack scenarios in synthetic microdata, where an adversary trains a predictive model on the synthetic dataset and attempts to infer sensitive attributes of real individuals using their quasi-identifiers. RAPID is the first bounded, interpretable risk measure that aligns with attacker capabilities, is independent of the synthesis mechanism, compatible with any learning algorithm, and robust to class imbalance. It handles continuous attributes via relative error tolerance, evaluates categorical attributes through baseline-normalized confidence scores, and integrates threshold calibration with uncertainty quantification. Experimental results demonstrate that RAPID provides tight, practical upper bounds on disclosure risk, effectively complementing existing privacy–utility evaluation frameworks.

0 citationsRead paper

Recent publications

Latest Papers

Agentic AI-Powered Re-Identification: An Emerging, Scalable Threat to Mobility Microdata Privacy

Jun 26, 2026

This work addresses the severe re-identification privacy risks posed by fine-grained location data collected by commercial data brokers, which traditional attacks struggle to scale due to reliance on manual analysis. The paper proposes the first end-to-end automated re-identification framework, leveraging large language model agents to autonomously harvest publicly available online information and integrate public records, social media profiles, and spatiotemporal trajectory matching algorithms—enabling large-scale identity inference without human intervention. Evaluated on a simulated dataset containing home and workplace address anchors, the method successfully re-identifies 18 out of 43 individuals (41.9%), achieving 72% accuracy among identifiable subjects. This study demonstrates, for the first time in a realistic setting, the feasibility of fully automated, low-cost, and highly efficient re-identification from mobile microdata, fundamentally challenging the conventional paradigm that depends on expert involvement.

0 citationsRead paper

Log-Ratio Propagation on the Simplex: A Theory of Cellwise Contamination for Compositional Data

May 29, 2026

This study addresses the ill-posedness of conventional Euclidean robust methods on the simplex when a single compositional component is contaminated, which induces a global shift in log-ratio coordinates. Building upon a scale-invariant multiplicative perturbation model, the work establishes the first theoretical framework for cellwise contamination in compositional data, proving that contamination in one component manifests as a rank-one shift in log-ratio space. The authors introduce a contamination propagation theorem and an influence function–based diagnostic fingerprint. By leveraging centered log-ratio transformation, isometric log-ratio coordinates, and contrast matrix analysis, they quantify the cellwise breakdown points of MCD, S-, τ-, and coordinatewise M-estimators, revealing a reduction by a factor of $(D-1)/D$ compared to their Euclidean counterparts. Furthermore, they demonstrate that the variability matrix’s influence function precisely identifies contaminated components, thereby laying the theoretical foundation for cellwise robust analysis of compositional data.

0 citationsRead paper

Low-Magnification SEM May Suffice: Interpretable Deep Learning for Multi-Scale Fracture-Cause Classification in Zirconia-Toughened Alumina

May 28, 2026

This study addresses the limitations of conventional fracture analysis for ceramic implants, which relies on high-magnification scanning electron microscopy (SEM)—a time-consuming and subjective process. The authors propose an interpretable deep learning approach based on Vision Transformers to automatically classify fracture origins in zirconia-toughened alumina ceramics—specifically green-state, hard-machining, and material defects—using multi-scale SEM images. They demonstrate for the first time that low-magnification SEM (50×) contains sufficient diagnostic information, achieving classification performance comparable to that of high-magnification images. By integrating Grad-CAM, the model provides spatially interpretable predictions aligned with established fractographic standards. Despite severe class imbalance, the method attains 0.907 accuracy and 0.888 macro F1-score, enabling effective low-magnification prescreening and substantially reducing reliance on high-magnification SEM.

0 citationsRead paper

Synthetic Image Detection with CLIP: Understanding and Assessing Predictive Cues

Feb 12, 2026

Recent generative models produce near-photorealistic images, challenging the trustworthiness of photographs. Synthetic image detection (SID) has thus become an important area of research. Prior work has highlighted how synthetic images differ from real photographs--unfortunately, SID methods often struggle to generalize to novel generative models and often perform poorly in practical settings. CLIP, a foundational vision-language model which yields semantically rich image-text embeddings, shows strong accuracy and generalization for SID. Yet, the underlying relevant cues embedded in CLIP-features remain unknown. It is unclear, whether CLIP-based detectors simply detect strong visual artifacts or exploit subtle semantic biases, both of which would render them useless in practical settings or on generative models of high quality. We introduce SynthCLIC, a paired dataset of real photographs and high-quality synthetic counterparts from recent diffusion models, designed to reduce semantic bias in SID. Using an interpretable linear head with de-correlated activations and a text-grounded concept-model, we analyze what CLIP-based detectors learn. CLIP-based linear detectors reach 0.96 mAP on a GAN-based benchmark but only 0.92 on our high-quality diffusion dataset SynthCLIC, and generalization across generator families drops to as low as 0.37 mAP. We find that the detectors primarily rely on high-level photographic attributes (e.g., minimalist style, lens flare, or depth layering), rather than overt generator-specific artifacts. CLIP-based detectors perform well overall but generalize unevenly across diverse generative architectures. This highlights the need for continual model updates and broader training exposure, while reinforcing CLIP-based approaches as a strong foundation for more universal, robust SID.

0 citationsRead paper

RAPID: Risk of Attribute Prediction-Induced Disclosure in Synthetic Microdata

Feb 09, 2026

This work proposes RAPID, a novel attribute inference risk metric designed to better reflect realistic attack scenarios in synthetic microdata, where an adversary trains a predictive model on the synthetic dataset and attempts to infer sensitive attributes of real individuals using their quasi-identifiers. RAPID is the first bounded, interpretable risk measure that aligns with attacker capabilities, is independent of the synthesis mechanism, compatible with any learning algorithm, and robust to class imbalance. It handles continuous attributes via relative error tolerance, evaluates categorical attributes through baseline-normalized confidence scores, and integrates threshold calibration with uncertainty quantification. Experimental results demonstrate that RAPID provides tight, practical upper bounds on disclosure risk, effectively complementing existing privacy–utility evaluation frameworks.

0 citationsRead paper