Selected work

Representative Papers

Picking the Right Image to Classify: Reliable-Input Selection in Teledermatology

Aug 17, 2026

This study addresses misclassification in teledermatology caused by distribution shifts and formally defines the Reliable Input Selection task with a comprehensive benchmark. By integrating training-free metrics—including embedding norms, neighborhood consensus, and prediction confidence—the proposed method selects optimal images from candidate sets to enhance the diagnostic performance of frozen models. Empirical results reveal a substantial gap between current approaches and the theoretical upper bound; while an ideal selector improves the weighted F1-score by approximately 20%, existing training-free strategies recover only partial gains. This work establishes a novel paradigm for mitigating acquisition discrepancies in remote dermatological diagnosis and identifies critical directions for future optimization in reliable input selection under domain shift conditions.

0 citationsRead paper

Consensus Measures for Unstructured Biomedical Text Annotations

Aug 04, 2026

This study addresses the challenge of quantifying soft inter-annotator agreement in open-label biomedical text annotation, where traditional metrics fall short due to the semantic flexibility of labels. To tackle this issue, the authors propose a semantic equivalence measure grounded in natural language inference (NLI), which effectively balances scalability with fine-grained conceptual discrimination. By integrating word embeddings, large language models, and NLI techniques, the method captures semantic similarity among open-ended labels in unstructured text. Synthetic experiments demonstrate that multiple semantic metrics can quantify soft agreement, yet they exhibit substantial differences in estimation bias. These findings underscore the novelty and practical utility of the proposed approach for evaluating annotation quality in complex, open-label settings.

0 citationsRead paper

Applying JEPA-Style Predictive Learning to JA4-Derived Network Fingerprints

Jul 09, 2026

This study introduces JEPA-style predictive learning to network traffic fingerprinting for the first time, addressing the challenge of learning generalizable representations under incomplete views across heterogeneous data sources. To this end, the authors propose JA4-JEPA, a Transformer-based model that leverages JA4-derived subfields—JA4, HA4H, JA4S, and JA4X—for self-supervised training. The model learns robust embeddings by enforcing consistency between predicted latent representations and outputs from a target encoder. Evaluated on 39,416 held-out samples, JA4-JEPA achieves a cosine similarity of 0.9899 and a kNN classification accuracy of 0.9220, demonstrating its capacity to extract high-quality network fingerprint representations even when input views are partially missing.

0 citationsRead paper

Convolutional-Neural-Networks for Deanonymisation of I2P Traffic

May 12, 2026

This study investigates the feasibility of de-anonymizing users in the I2P network through passive traffic analysis. To address potential unique patterns in encrypted traffic, the authors generate synthetic I2P traffic in a controlled environment and, for the first time, integrate convolutional neural networks with Fano’s inequality from information theory to data-drivenly uncover causal relationships among traffic flows and assess the de-anonymization capability of deep learning models on real I2P traffic. Experimental results demonstrate that the proposed approach fails to breach the anonymity guarantees provided by I2P under its current configuration, thereby empirically validating the effectiveness of its anonymity mechanisms. Furthermore, this work introduces a novel methodology for analyzing the theoretical limits of anonymous communication networks.

0 citationsRead paper

Explainable Part-Based Vehicle Classifier with Spatial Awareness

May 08, 2026

This work addresses the limited interpretability and poor robustness to part misdetection in fine-grained vehicle classification models for intelligent transportation systems. To overcome these challenges, the authors propose a decoupled architecture that decomposes an end-to-end CNN into three components: a semantic part detector, a spatial-aware feature construction module, and a softmax regression classifier. Innovatively replacing the conventional binary part-presence indicator with spatial probability maps of vehicle parts significantly enhances the model’s robustness against part misdetection. The proposed approach achieves classification accuracy comparable to state-of-the-art end-to-end CNNs while substantially improving interpretability, thereby breaking the traditional trade-off between accuracy and explainability.

0 citationsRead paper

Recent publications

Latest Papers

Picking the Right Image to Classify: Reliable-Input Selection in Teledermatology

Aug 17, 2026

This study addresses misclassification in teledermatology caused by distribution shifts and formally defines the Reliable Input Selection task with a comprehensive benchmark. By integrating training-free metrics—including embedding norms, neighborhood consensus, and prediction confidence—the proposed method selects optimal images from candidate sets to enhance the diagnostic performance of frozen models. Empirical results reveal a substantial gap between current approaches and the theoretical upper bound; while an ideal selector improves the weighted F1-score by approximately 20%, existing training-free strategies recover only partial gains. This work establishes a novel paradigm for mitigating acquisition discrepancies in remote dermatological diagnosis and identifies critical directions for future optimization in reliable input selection under domain shift conditions.

0 citationsRead paper

Consensus Measures for Unstructured Biomedical Text Annotations

Aug 04, 2026

This study addresses the challenge of quantifying soft inter-annotator agreement in open-label biomedical text annotation, where traditional metrics fall short due to the semantic flexibility of labels. To tackle this issue, the authors propose a semantic equivalence measure grounded in natural language inference (NLI), which effectively balances scalability with fine-grained conceptual discrimination. By integrating word embeddings, large language models, and NLI techniques, the method captures semantic similarity among open-ended labels in unstructured text. Synthetic experiments demonstrate that multiple semantic metrics can quantify soft agreement, yet they exhibit substantial differences in estimation bias. These findings underscore the novelty and practical utility of the proposed approach for evaluating annotation quality in complex, open-label settings.

0 citationsRead paper

Applying JEPA-Style Predictive Learning to JA4-Derived Network Fingerprints

Jul 09, 2026

This study introduces JEPA-style predictive learning to network traffic fingerprinting for the first time, addressing the challenge of learning generalizable representations under incomplete views across heterogeneous data sources. To this end, the authors propose JA4-JEPA, a Transformer-based model that leverages JA4-derived subfields—JA4, HA4H, JA4S, and JA4X—for self-supervised training. The model learns robust embeddings by enforcing consistency between predicted latent representations and outputs from a target encoder. Evaluated on 39,416 held-out samples, JA4-JEPA achieves a cosine similarity of 0.9899 and a kNN classification accuracy of 0.9220, demonstrating its capacity to extract high-quality network fingerprint representations even when input views are partially missing.

0 citationsRead paper

Convolutional-Neural-Networks for Deanonymisation of I2P Traffic

May 12, 2026

This study investigates the feasibility of de-anonymizing users in the I2P network through passive traffic analysis. To address potential unique patterns in encrypted traffic, the authors generate synthetic I2P traffic in a controlled environment and, for the first time, integrate convolutional neural networks with Fano’s inequality from information theory to data-drivenly uncover causal relationships among traffic flows and assess the de-anonymization capability of deep learning models on real I2P traffic. Experimental results demonstrate that the proposed approach fails to breach the anonymity guarantees provided by I2P under its current configuration, thereby empirically validating the effectiveness of its anonymity mechanisms. Furthermore, this work introduces a novel methodology for analyzing the theoretical limits of anonymous communication networks.

0 citationsRead paper

Explainable Part-Based Vehicle Classifier with Spatial Awareness

May 08, 2026

This work addresses the limited interpretability and poor robustness to part misdetection in fine-grained vehicle classification models for intelligent transportation systems. To overcome these challenges, the authors propose a decoupled architecture that decomposes an end-to-end CNN into three components: a semantic part detector, a spatial-aware feature construction module, and a softmax regression classifier. Innovatively replacing the conventional binary part-presence indicator with spatial probability maps of vehicle parts significantly enhances the model’s robustness against part misdetection. The proposed approach achieves classification accuracy comparable to state-of-the-art end-to-end CNNs while substantially improving interpretability, thereby breaking the traditional trade-off between accuracy and explainability.

0 citationsRead paper