Dongxiao Zhu
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Dongxiao Zhu has a total of 98 co-authors. Here are the top co-authors by number of shared publications:
- Xiangyu Zhou
Wayne State University, Computer science, Detroit, Michigan, USA
7 shared publications Yao Qiang
Wayne State University
7 shared publicationsXin Li
Wayne State University
6 shared publications- Yao Qiang
School of Engineering and Computer Science, Oakland University, Rochester, Michigan, USA
6 shared publications - Chengyin Li
Wayne State University
5 shared publications - Deng Pan
Wayne State University
5 shared publications
Latest Publications
- 5 September 2026
- Published by Springer Nature
- p. 224-241
Hijacking Large Language Models via Adversarial In-Context Learning
- 5 September 2026
- Published by Springer Nature
- p. 224-241
Hijacking Large Language Models via Adversarial In-Context Learning
- Open Access
- 13 June 2026
- Published by Association for Computing Machinery (ACM)
- p. 172-191
Trauma-Informed Data Donation: Integrating Expert and Donor Perspectives on Designing Against Re-Traumatization During Collection of Sexual Violence Data
- 13 June 2026
- Published by Association for Computing Machinery (ACM)
- p. 172-191
Open AccessTrauma-Informed Data Donation: Integrating Expert and Donor Perspectives on Designing Against Re-Traumatization During Collection of Sexual Violence Data
- Open Access
- 12 March 2026
- Published by JMIR Publications Inc. in JMIR AI
- Vol. 5  (1) , e91981
Authors’ Reply: Toward Retrieval-Grounded Evaluation for Conversational Large Language Model–Based Risk Assessment
- 12 March 2026
- Published by JMIR Publications Inc. in JMIR AI
- Vol. 5  (1) , e91981
Open AccessAuthors’ Reply: Toward Retrieval-Grounded Evaluation for Conversational Large Language Model–Based Risk Assessment
- Open Access
- 16 October 2025
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Human-Computer Interaction
- Vol. 9  (7) , 1-30
Collective Consent: Who Needs to Consent to the Donation of Data Representing Multiple People?
Data donation is a growing form of personal data collection that foregrounds consent and conscious participation of the data donor. There remains little guidance on who must consent to data donation, particularly when the data represents multiple people. We provide empirical perspectives on this question through in-situ observation and interviews (N=18) with online daters who chose to donate messaging interactions with potential sexual partners for sexual violence research. Findings elucidate two diverging perspectives. Participants advocating for ''unilateral consent'' argued that consent of their messaging partner is not necessary, in part, because the anticipated benefit of data donation superseded consent. Participants advocating for ''collective consent'' wanted both messaging partners to consent to its donation, citing concerns for privacy of, and personal relationships with, the other person. Findings suggest that collective consent interfaces should be incorporated in data donation platforms, even if not strictly required by legal regulation, to improve donation of multi-person data.
- 16 October 2025
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Human-Computer Interaction
- Vol. 9  (7) , 1-30
Open AccessCollective Consent: Who Needs to Consent to the Donation of Data Representing Multiple People?
Data donation is a growing form of personal data collection that foregrounds consent and conscious participation of the data donor. There remains little guidance on who must consent to data donation, particularly when the data represents multiple people. We provide empirical perspectives on this question through in-situ observation and interviews (N=18) with online daters who chose to donate messaging interactions with potential sexual partners for sexual violence research. Findings elucidate two diverging perspectives. Participants advocating for ''unilateral consent'' argued that consent of their messaging partner is not necessary, in part, because the anticipated benefit of data donation superseded consent. Participants advocating for ''collective consent'' wanted both messaging partners to consent to its donation, citing concerns for privacy of, and personal relationships with, the other person. Findings suggest that collective consent interfaces should be incorporated in data donation platforms, even if not strictly required by legal regulation, to improve donation of multi-person data.
- Open Access
Automatic Calibration for Membership Inference Attack on Large Language Models
Membership Inference Attacks (MIAs) have recently been employed to determine whether a specific text was part of the pre-training data of Large Language Models (LLMs). However, existing methods often misinfer non-members as members, leading to a high false positive rate, or depend on additional reference models for probability calibration, which limits their practicality. To overcome these challenges, we introduce a novel framework called Automatic Calibration Membership Inference Attack (ACMIA), which utilizes a tunable temperature to calibrate output probabilities effectively. This approach is inspired by our theoretical insights into maximum likelihood estimation during the pre-training of LLMs. We introduce ACMIA in three configurations designed to accommodate different levels of model access and increase the probability gap between members and non-members, improving the reliability and robustness of membership inference. Extensive experiments on various open-source LLMs demonstrate that our proposed attack is highly effective, robust, and generalizable, surpassing state-of-the-art baselines across three widely used benchmarks. Our code is available at: \href{https://github.com/Salehzz/ACMIA}{\textcolor{blue}{Github}}.
Open AccessAutomatic Calibration for Membership Inference Attack on Large Language Models
Membership Inference Attacks (MIAs) have recently been employed to determine whether a specific text was part of the pre-training data of Large Language Models (LLMs). However, existing methods often misinfer non-members as members, leading to a high false positive rate, or depend on additional reference models for probability calibration, which limits their practicality. To overcome these challenges, we introduce a novel framework called Automatic Calibration Membership Inference Attack (ACMIA), which utilizes a tunable temperature to calibrate output probabilities effectively. This approach is inspired by our theoretical insights into maximum likelihood estimation during the pre-training of LLMs. We introduce ACMIA in three configurations designed to accommodate different levels of model access and increase the probability gap between members and non-members, improving the reliability and robustness of membership inference. Extensive experiments on various open-source LLMs demonstrate that our proposed attack is highly effective, robust, and generalizable, surpassing state-of-the-art baselines across three widely used benchmarks. Our code is available at: \href{https://github.com/Salehzz/ACMIA}{\textcolor{blue}{Github}}.
- Open Access
BiPVL-Seg: Bidirectional Progressive Vision-Language Fusion with Global-Local Alignment for Medical Image Segmentation
Medical image segmentation typically relies solely on visual data, overlooking the rich textual information clinicians use for diagnosis. Vision-language models attempt to bridge this gap, but existing approaches often process visual and textual features independently, resulting in weak cross-modal alignment. Simple fusion techniques fail due to the inherent differences between spatial visual features and sequential text embeddings. Additionally, medical terminology deviates from general language, limiting the effectiveness of off-the-shelf text encoders and further hindering vision-language alignment. We propose BiPVL-Seg, an end-to-end framework that integrates vision-language fusion and embedding alignment through architectural and training innovations, where both components reinforce each other to enhance medical image segmentation. BiPVL-Seg introduces bidirectional progressive fusion in the architecture, which facilitates stage-wise information exchange between vision and text encoders. Additionally, it incorporates global-local contrastive alignment, a training objective that enhances the text encoder's comprehension by aligning text and vision embeddings at both class and concept levels. Extensive experiments on diverse medical imaging benchmarks across CT and MR modalities demonstrate BiPVL-Seg's superior performance when compared with state-of-the-art methods in complex multi-class segmentation. Source code is available in this GitHub repository.
Open AccessBiPVL-Seg: Bidirectional Progressive Vision-Language Fusion with Global-Local Alignment for Medical Image Segmentation
Medical image segmentation typically relies solely on visual data, overlooking the rich textual information clinicians use for diagnosis. Vision-language models attempt to bridge this gap, but existing approaches often process visual and textual features independently, resulting in weak cross-modal alignment. Simple fusion techniques fail due to the inherent differences between spatial visual features and sequential text embeddings. Additionally, medical terminology deviates from general language, limiting the effectiveness of off-the-shelf text encoders and further hindering vision-language alignment. We propose BiPVL-Seg, an end-to-end framework that integrates vision-language fusion and embedding alignment through architectural and training innovations, where both components reinforce each other to enhance medical image segmentation. BiPVL-Seg introduces bidirectional progressive fusion in the architecture, which facilitates stage-wise information exchange between vision and text encoders. Additionally, it incorporates global-local contrastive alignment, a training objective that enhances the text encoder's comprehension by aligning text and vision embeddings at both class and concept levels. Extensive experiments on diverse medical imaging benchmarks across CT and MR modalities demonstrate BiPVL-Seg's superior performance when compared with state-of-the-art methods in complex multi-class segmentation. Source code is available in this GitHub repository.
- Open Access
- 27 March 2025
- Published by JMIR Publications Inc. in JMIR AI
- Vol. 4, e67363
Generative Large Language Model—Powered Conversational AI App for Personalized Risk Assessment: Case Study in COVID-19
Background: Large language models (LLMs) have demonstrated powerful capabilities in natural language tasks and are increasingly being integrated into health care for tasks like disease risk assessment. Traditional machine learning methods rely on structured data and coding, limiting their flexibility in dynamic clinical environments. This study presents a novel approach to disease risk assessment using generative LLMs through conversational artificial intelligence (AI), eliminating the need for programming. Objective: This study evaluates the use of pretrained generative LLMs, including LLaMA2-7b and Flan-T5-xl, for COVID-19 severity prediction with the goal of enabling a real-time, no-code, risk assessment solution through chatbot-based, question-answering interactions. To contextualize their performance, we compare LLMs with traditional machine learning classifiers, such as logistic regression, extreme gradient boosting (XGBoost), and random forest, which rely on tabular data. Methods: We fine-tuned LLMs using few-shot natural language examples from a dataset of 393 pediatric patients, developing a mobile app that integrates these models to provide real-time, no-code, COVID-19 severity risk assessment through clinician-patient interaction. The LLMs were compared with traditional classifiers across different experimental settings, using the area under the curve (AUC) as the primary evaluation metric. Feature importance derived from LLM attention layers was also analyzed to enhance interpretability. Results: Generative LLMs demonstrated strong performance in low-data settings. In zero-shot scenarios, the T0-3b-T model achieved an AUC of 0.75, while other LLMs, such as T0pp(8bit)-T and Flan-T5-xl-T, reached 0.67 and 0.69, respectively. At 2-shot settings, logistic regression and random forest achieved an AUC of 0.57, while Flan-T5-xl-T and T0-3b-T obtained 0.69 and 0.65, respectively. By 32-shot settings, Flan-T5-xl-T reached 0.70, similar to logistic regression (0.69) and random forest (0.68), while XGBoost improved to 0.65. These results illustrate the differences in how generative LLMs and traditional models handle the increasing data availability. LLMs perform well in low-data scenarios, whereas traditional models rely more on structured tabular data and labeled training examples. Furthermore, the mobile app provides real-time, COVID-19 severity assessments and personalized insights through attention-based feature importance, adding value to the clinical interpretation of the results. Conclusions: Generative LLMs provide a robust alternative to traditional classifiers, particularly in scenarios with limited labeled data. Their ability to handle unstructured inputs and deliver personalized, real-time assessments without coding makes them highly adaptable to clinical settings. This study underscores the potential of LLM-powered conversational artificial intelligence (AI) in health care and encourages further exploration of its use for real-time, disease risk assessment and decision-making support.
- 27 March 2025
- Published by JMIR Publications Inc. in JMIR AI
- Vol. 4, e67363
Open AccessGenerative Large Language Model—Powered Conversational AI App for Personalized Risk Assessment: Case Study in COVID-19
Background: Large language models (LLMs) have demonstrated powerful capabilities in natural language tasks and are increasingly being integrated into health care for tasks like disease risk assessment. Traditional machine learning methods rely on structured data and coding, limiting their flexibility in dynamic clinical environments. This study presents a novel approach to disease risk assessment using generative LLMs through conversational artificial intelligence (AI), eliminating the need for programming. Objective: This study evaluates the use of pretrained generative LLMs, including LLaMA2-7b and Flan-T5-xl, for COVID-19 severity prediction with the goal of enabling a real-time, no-code, risk assessment solution through chatbot-based, question-answering interactions. To contextualize their performance, we compare LLMs with traditional machine learning classifiers, such as logistic regression, extreme gradient boosting (XGBoost), and random forest, which rely on tabular data. Methods: We fine-tuned LLMs using few-shot natural language examples from a dataset of 393 pediatric patients, developing a mobile app that integrates these models to provide real-time, no-code, COVID-19 severity risk assessment through clinician-patient interaction. The LLMs were compared with traditional classifiers across different experimental settings, using the area under the curve (AUC) as the primary evaluation metric. Feature importance derived from LLM attention layers was also analyzed to enhance interpretability. Results: Generative LLMs demonstrated strong performance in low-data settings. In zero-shot scenarios, the T0-3b-T model achieved an AUC of 0.75, while other LLMs, such as T0pp(8bit)-T and Flan-T5-xl-T, reached 0.67 and 0.69, respectively. At 2-shot settings, logistic regression and random forest achieved an AUC of 0.57, while Flan-T5-xl-T and T0-3b-T obtained 0.69 and 0.65, respectively. By 32-shot settings, Flan-T5-xl-T reached 0.70, similar to logistic regression (0.69) and random forest (0.68), while XGBoost improved to 0.65. These results illustrate the differences in how generative LLMs and traditional models handle the increasing data availability. LLMs perform well in low-data scenarios, whereas traditional models rely more on structured tabular data and labeled training examples. Furthermore, the mobile app provides real-time, COVID-19 severity assessments and personalized insights through attention-based feature importance, adding value to the clinical interpretation of the results. Conclusions: Generative LLMs provide a robust alternative to traditional classifiers, particularly in scenarios with limited labeled data. Their ability to handle unstructured inputs and deliver personalized, real-time assessments without coding makes them highly adaptable to clinical settings. This study underscores the potential of LLM-powered conversational artificial intelligence (AI) in health care and encourages further exploration of its use for real-time, disease risk assessment and decision-making support.
- 2 December 2024
- Published by Springer Nature
- p. 358-376
Fairness-Aware Vision Transformer via Debiased Self-Attention
- 2 December 2024
- Published by Springer Nature
- p. 358-376
Fairness-Aware Vision Transformer via Debiased Self-Attention
- Open Access
MulModSeg: Enhancing Unpaired Multi-Modal Medical Image Segmentation with Modality-Conditioned Text Embedding and Alternating Training
In the diverse field of medical imaging, automatic segmentation has numerous applications and must handle a wide variety of input domains, such as different types of Computed Tomography (CT) scans and Magnetic Resonance (MR) images. This heterogeneity challenges automatic segmentation algorithms to maintain consistent performance across different modalities due to the requirement for spatially aligned and paired images. Typically, segmentation models are trained using a single modality, which limits their ability to generalize to other types of input data without employing transfer learning techniques. Additionally, leveraging complementary information from different modalities to enhance segmentation precision often necessitates substantial modifications to popular encoder-decoder designs, such as introducing multiple branched encoding or decoding paths for each modality. In this work, we propose a simple Multi-Modal Segmentation (MulModSeg) strategy to enhance medical image segmentation across multiple modalities, specifically CT and MR. It incorporates two key designs: a modality-conditioned text embedding framework via a frozen text encoder that adds modality awareness to existing segmentation frameworks without significant structural modifications or computational overhead, and an alternating training procedure that facilitates the integration of essential features from unpaired CT and MR inputs. Through extensive experiments with both Fully Convolutional Network and Transformer-based backbones, MulModSeg consistently outperforms previous methods in segmenting abdominal multi-organ and cardiac substructures for both CT and MR modalities. The code is available in this {\href{https://github.com/ChengyinLee/MulModSeg_2024}{link}}.
Open AccessMulModSeg: Enhancing Unpaired Multi-Modal Medical Image Segmentation with Modality-Conditioned Text Embedding and Alternating Training
In the diverse field of medical imaging, automatic segmentation has numerous applications and must handle a wide variety of input domains, such as different types of Computed Tomography (CT) scans and Magnetic Resonance (MR) images. This heterogeneity challenges automatic segmentation algorithms to maintain consistent performance across different modalities due to the requirement for spatially aligned and paired images. Typically, segmentation models are trained using a single modality, which limits their ability to generalize to other types of input data without employing transfer learning techniques. Additionally, leveraging complementary information from different modalities to enhance segmentation precision often necessitates substantial modifications to popular encoder-decoder designs, such as introducing multiple branched encoding or decoding paths for each modality. In this work, we propose a simple Multi-Modal Segmentation (MulModSeg) strategy to enhance medical image segmentation across multiple modalities, specifically CT and MR. It incorporates two key designs: a modality-conditioned text embedding framework via a frozen text encoder that adds modality awareness to existing segmentation frameworks without significant structural modifications or computational overhead, and an alternating training procedure that facilitates the integration of essential features from unpaired CT and MR inputs. Through extensive experiments with both Fully Convolutional Network and Transformer-based backbones, MulModSeg consistently outperforms previous methods in segmenting abdominal multi-organ and cardiac substructures for both CT and MR modalities. The code is available in this {\href{https://github.com/ChengyinLee/MulModSeg_2024}{link}}.
- Open Access
- 9 October 2024
- Published by JMIR Publications Inc. in JMIR Preprints
Generative Large Language Model—Powered Conversational AI App for Personalized Risk Assessment: Case Study in COVID-19 (Preprint)
BACKGROUND Large language models (LLMs) have demonstrated powerful capabilities in natural language tasks and are increasingly being integrated into health care for tasks like disease risk assessment. Traditional machine learning methods rely on structured data and coding, limiting their flexibility in dynamic clinical environments. This study presents a novel approach to disease risk assessment using generative LLMs through conversational artificial intelligence (AI), eliminating the need for programming. OBJECTIVE This study evaluates the use of pretrained generative LLMs, including LLaMA2-7b and Flan-T5-xl, for COVID-19 severity prediction with the goal of enabling a real-time, no-code, risk assessment solution through chatbot-based, question-answering interactions. To contextualize their performance, we compare LLMs with traditional machine learning classifiers, such as logistic regression, extreme gradient boosting (XGBoost), and random forest, which rely on tabular data. METHODS We fine-tuned LLMs using few-shot natural language examples from a dataset of 393 pediatric patients, developing a mobile app that integrates these models to provide real-time, no-code, COVID-19 severity risk assessment through clinician-patient interaction. The LLMs were compared with traditional classifiers across different experimental settings, using the area under the curve (AUC) as the primary evaluation metric. Feature importance derived from LLM attention layers was also analyzed to enhance interpretability. RESULTS Generative LLMs demonstrated strong performance in low-data settings. In zero-shot scenarios, the T0-3b-T model achieved an AUC of 0.75, while other LLMs, such as T0pp(8bit)-T and Flan-T5-xl-T, reached 0.67 and 0.69, respectively. At 2-shot settings, logistic regression and random forest achieved an AUC of 0.57, while Flan-T5-xl-T and T0-3b-T obtained 0.69 and 0.65, respectively. By 32-shot settings, Flan-T5-xl-T reached 0.70, similar to logistic regression (0.69) and random forest (0.68), while XGBoost improved to 0.65. These results illustrate the differences in how generative LLMs and traditional models handle the increasing data availability. LLMs perform well in low-data scenarios, whereas traditional models rely more on structured tabular data and labeled training examples. Furthermore, the mobile app provides real-time, COVID-19 severity assessments and personalized insights through attention-based feature importance, adding value to the clinical interpretation of the results. CONCLUSIONS Generative LLMs provide a robust alternative to traditional classifiers, particularly in scenarios with limited labeled data. Their ability to handle unstructured inputs and deliver personalized, real-time assessments without coding makes them highly adaptable to clinical settings. This study underscores the potential of LLM-powered conversational artificial intelligence (AI) in health care and encourages further exploration of its use for real-time, disease risk assessment and decision-making support.
- 9 October 2024
- Published by JMIR Publications Inc. in JMIR Preprints
Open AccessGenerative Large Language Model—Powered Conversational AI App for Personalized Risk Assessment: Case Study in COVID-19 (Preprint)
BACKGROUND Large language models (LLMs) have demonstrated powerful capabilities in natural language tasks and are increasingly being integrated into health care for tasks like disease risk assessment. Traditional machine learning methods rely on structured data and coding, limiting their flexibility in dynamic clinical environments. This study presents a novel approach to disease risk assessment using generative LLMs through conversational artificial intelligence (AI), eliminating the need for programming. OBJECTIVE This study evaluates the use of pretrained generative LLMs, including LLaMA2-7b and Flan-T5-xl, for COVID-19 severity prediction with the goal of enabling a real-time, no-code, risk assessment solution through chatbot-based, question-answering interactions. To contextualize their performance, we compare LLMs with traditional machine learning classifiers, such as logistic regression, extreme gradient boosting (XGBoost), and random forest, which rely on tabular data. METHODS We fine-tuned LLMs using few-shot natural language examples from a dataset of 393 pediatric patients, developing a mobile app that integrates these models to provide real-time, no-code, COVID-19 severity risk assessment through clinician-patient interaction. The LLMs were compared with traditional classifiers across different experimental settings, using the area under the curve (AUC) as the primary evaluation metric. Feature importance derived from LLM attention layers was also analyzed to enhance interpretability. RESULTS Generative LLMs demonstrated strong performance in low-data settings. In zero-shot scenarios, the T0-3b-T model achieved an AUC of 0.75, while other LLMs, such as T0pp(8bit)-T and Flan-T5-xl-T, reached 0.67 and 0.69, respectively. At 2-shot settings, logistic regression and random forest achieved an AUC of 0.57, while Flan-T5-xl-T and T0-3b-T obtained 0.69 and 0.65, respectively. By 32-shot settings, Flan-T5-xl-T reached 0.70, similar to logistic regression (0.69) and random forest (0.68), while XGBoost improved to 0.65. These results illustrate the differences in how generative LLMs and traditional models handle the increasing data availability. LLMs perform well in low-data scenarios, whereas traditional models rely more on structured tabular data and labeled training examples. Furthermore, the mobile app provides real-time, COVID-19 severity assessments and personalized insights through attention-based feature importance, adding value to the clinical interpretation of the results. CONCLUSIONS Generative LLMs provide a robust alternative to traditional classifiers, particularly in scenarios with limited labeled data. Their ability to handle unstructured inputs and deliver personalized, real-time assessments without coding makes them highly adaptable to clinical settings. This study underscores the potential of LLM-powered conversational artificial intelligence (AI) in health care and encourages further exploration of its use for real-time, disease risk assessment and decision-making support.