Institution profile

University of New Brunswick

Academic institutionnorthamerica · ca
Official website
Research library95linked papers
Opportunities0open roles
Selected work

Representative Papers

GraphRectify: Graph-Based Transfer of Adversarial Example Detectors Across Neural Networks

Oct 07, 2026

This study addresses the limited cross-model reusability of existing adversarial example detectors caused by their dependence on specific classification backbones. We propose a graph-based universal detection framework that leverages graph neural networks to transform intermediate-layer features into structured representations. By introducing a cross-backbone representation alignment technique, our approach overcomes the bottleneck of ineffective transfer of detection knowledge across heterogeneous architectures, enabling direct detector reuse without training from scratch. Extensive evaluations across multiple datasets and adaptive attack scenarios demonstrate that the proposed method achieves significantly superior aggregated ROC-AUC compared to both from-scratch training and existing transfer baselines. This work establishes a novel paradigm for constructing plug-and-play, highly generalizable adversarial defense systems.

0 citationsRead paper

Detecting Adversarial Images through Response Profiles of Vision-Language Models

Oct 07, 2026

This study addresses the vulnerability of frozen vision-language models to adversarial image perturbations and the inadequacy of existing detection mechanisms. It proposes the first detection paradigm based on deviations in multi-prompt response distributions and their stability. Using CLIP as the backbone, the method generates responses through diverse semantic prompts, extracts statistical features to construct compact profiles, and trains a lightweight classifier to identify anomalous inputs. The proposed approach achieves strong discriminative capability across multiple adversarial attacks, significantly outperforming embedding-space geometric baselines while demonstrating superior cross-attack generalization. Ultimately, this work provides an efficient and reliable defense strategy for the secure deployment of frozen vision-language models.

0 citationsRead paper

ExpertoRhythm: Morphology-Aware Learning for Waveform Reconstruction and Cuffless Blood Pressure Estimation from Single-Channel PPG

Sep 28, 2026Conference on Artificial Intelligence in Medicine in Europe

This study addresses the challenge of preserving waveform morphology and dynamic variations when estimating blood pressure from single-channel photoplethysmography (PPG). We propose a cuffless continuous blood pressure monitoring method based on an attention-enhanced 1D U-Net. By designing a composite morphology-aware objective function that combines range-weighted SmoothL1 loss with window-based regularization, the approach amplifies high-dynamic segment features while suppressing amplitude errors, enabling precise PPG-to-arterial blood pressure waveform reconstruction and direct systolic/diastolic blood pressure (SBP/DBP) estimation. Experimental results demonstrate mean absolute errors of 2.46 mmHg for SBP and 1.46 mmHg for DBP, achieving a 30.4% relative error reduction over the MSE baseline and satisfying both AAMI and BHS Grade A clinical standards.

0 citationsRead paper

CAMEO: A Class-Activation-Mapped Equitable Overlay Framework for Fair and Robust Deep Learning-based Skin Condition Diagnosis

Sep 28, 2026

This study addresses the vulnerability of deep learning-based dermatological diagnostic models to shortcut learning, wherein reliance on background artifacts such as skin color compromises fairness and robustness. To mitigate this issue, we propose CAMEO, a novel framework that introduces an explainable AI (XAI)-guided data augmentation mechanism. By leveraging class activation mapping, CAMEO achieves annotation-free lesion localization and background decoupling, subsequently eliminating spurious correlations through synthetic skin replacement. Evaluated on the HAM10000 dataset, the proposed method reduces background-driven errors by nearly fourfold while preserving classification accuracy. Furthermore, it significantly enhances cross-skin-tone generalization, effectively reconciling diagnostic precision with algorithmic fairness in automated dermatological assessment.

0 citationsRead paper
Recent publications

Latest Papers

GraphRectify: Graph-Based Transfer of Adversarial Example Detectors Across Neural Networks

Oct 07, 2026

This study addresses the limited cross-model reusability of existing adversarial example detectors caused by their dependence on specific classification backbones. We propose a graph-based universal detection framework that leverages graph neural networks to transform intermediate-layer features into structured representations. By introducing a cross-backbone representation alignment technique, our approach overcomes the bottleneck of ineffective transfer of detection knowledge across heterogeneous architectures, enabling direct detector reuse without training from scratch. Extensive evaluations across multiple datasets and adaptive attack scenarios demonstrate that the proposed method achieves significantly superior aggregated ROC-AUC compared to both from-scratch training and existing transfer baselines. This work establishes a novel paradigm for constructing plug-and-play, highly generalizable adversarial defense systems.

0 citationsRead paper

Detecting Adversarial Images through Response Profiles of Vision-Language Models

Oct 07, 2026

This study addresses the vulnerability of frozen vision-language models to adversarial image perturbations and the inadequacy of existing detection mechanisms. It proposes the first detection paradigm based on deviations in multi-prompt response distributions and their stability. Using CLIP as the backbone, the method generates responses through diverse semantic prompts, extracts statistical features to construct compact profiles, and trains a lightweight classifier to identify anomalous inputs. The proposed approach achieves strong discriminative capability across multiple adversarial attacks, significantly outperforming embedding-space geometric baselines while demonstrating superior cross-attack generalization. Ultimately, this work provides an efficient and reliable defense strategy for the secure deployment of frozen vision-language models.

0 citationsRead paper

ExpertoRhythm: Morphology-Aware Learning for Waveform Reconstruction and Cuffless Blood Pressure Estimation from Single-Channel PPG

Sep 28, 2026Conference on Artificial Intelligence in Medicine in Europe

This study addresses the challenge of preserving waveform morphology and dynamic variations when estimating blood pressure from single-channel photoplethysmography (PPG). We propose a cuffless continuous blood pressure monitoring method based on an attention-enhanced 1D U-Net. By designing a composite morphology-aware objective function that combines range-weighted SmoothL1 loss with window-based regularization, the approach amplifies high-dynamic segment features while suppressing amplitude errors, enabling precise PPG-to-arterial blood pressure waveform reconstruction and direct systolic/diastolic blood pressure (SBP/DBP) estimation. Experimental results demonstrate mean absolute errors of 2.46 mmHg for SBP and 1.46 mmHg for DBP, achieving a 30.4% relative error reduction over the MSE baseline and satisfying both AAMI and BHS Grade A clinical standards.

0 citationsRead paper

CAMEO: A Class-Activation-Mapped Equitable Overlay Framework for Fair and Robust Deep Learning-based Skin Condition Diagnosis

Sep 28, 2026

This study addresses the vulnerability of deep learning-based dermatological diagnostic models to shortcut learning, wherein reliance on background artifacts such as skin color compromises fairness and robustness. To mitigate this issue, we propose CAMEO, a novel framework that introduces an explainable AI (XAI)-guided data augmentation mechanism. By leveraging class activation mapping, CAMEO achieves annotation-free lesion localization and background decoupling, subsequently eliminating spurious correlations through synthetic skin replacement. Evaluated on the HAM10000 dataset, the proposed method reduces background-driven errors by nearly fourfold while preserving classification accuracy. Furthermore, it significantly enhances cross-skin-tone generalization, effectively reconciling diagnostic precision with algorithmic fairness in automated dermatological assessment.

0 citationsRead paper