Dr. Gongbo "Tony" Liang - Research and Projects

Research Overview

As an IEEE Senior Member and Trustworthy AI researcher, I have been advancing the field since 2017 by making neural networks more reliable, robust, and fair. My research in Trustworthy AI directly strengthens modern cybersecurity by tackling foundational challenges in calibration, bias mitigation in generative AI, uncertainty estimation, and explainability. I develop practical solutions for improving model robustness, including black-box defenses against adversarial attacks, and optimize training efficiency through innovative approaches to dataset reduction, pre-training, and cross-modality analysis. I apply these advancements to high-impact domains, including medical imaging, public health, transportation safety, and astrophysics, to ensure AI solutions are secure, effective, and trustworthy for real-world problems.

Foundational Research in Trustworthy Artificial Intelligence

Abstract: This research program moves beyond predictive accuracy to address the core principles of trustworthy AI. The goal is to develop a comprehensive theoretical framework for building artificial intelligence systems that are fundamentally reliable, fair, transparent, and efficient. Our work investigates several key thrusts: we enhance model reliability and security by improving robustness against adversarial attacks and developing techniques for uncertainty estimation, calibration, model stability, reproducibility, and effective domain adaptation. We address fairness and transparency by creating novel methods for bias mitigation and advancing the field of explainability (XAI). Finally, we improve training efficiency by exploring methods like contrastive learning to reduce the dependence on large-scale training data. The overarching vision of this project is to create foundational, domain-agnostic principles that ensure AI systems can be safely and ethically deployed in the real world.

Keywords: Calibration, Bias Mitigation, Uncertainty Estimation, Explainability (XAI), Dataset Reduction, Contrastive Learning, Model Stability, Model Reproducibility, Domain Adaptation, Adversarial Attacks

Selected Papers: Defense-PointNet (BigData'19), NN Calibration (ICML'20 Workshop), Dynamic Feature Alignment (BMVC'21), NN Feature Inconsistency (ICPR'22), LLM Adversarial Attacking (Electronics'23), Rust Vulnerability Detection (MLKE'25)

A comparison of probability calibration methods on a synthetic dataset. The diagonal line represents perfect calibration ('True Prob'). (Left) The uncalibrated model is poorly calibrated and exhibits high overconfidence. (Middle) Temperature scaling, a common technique, reduces overconfidence but fails to correct the underlying distribution. (Right) Our proposed method accurately recovers the true probability distribution across the input space.

Advancing Uncertainty-Aware GeoAI for Roadway Safety in Rural and Micropolitan USA

Abstract: The United States faces a critical socio-technical safety crisis, with rural and micropolitan regions bearing a disproportionate burden of traffic fatalities, accounting for 40% of deaths despite comprising only 19% of the population. The inherent uncertainty of traffic crashes, driven by complex interactions among driver behavior, environmental conditions, and infrastructure, renders traditional deterministic AI risk estimators inadequate, as they obscure confidence levels and undermine trust in safety decisions. This project proposes a novel system that integrates the physical roadway network with a GeoAI framework, designed to embrace uncertainty and deliver trustworthy safety analytics for resource-constrained rural areas. Our long-term vision introduces a Dynamic Geospatial Knowledge Engine (DGKE) to create dynamic, context-aware knowledge graphs. By incorporating causal inference, the DGKE aims to provide real-time, personalized safety insights, transforming traffic safety through a collaborative human-AI approach.

Keywords: Public Health, Multi-Modal, Uncertainty Estimation, Explainability, Calibration

Selected Papers: MTSL-RoadRisk (IEEE J-STARS'23), SA Case Study (AAAI'25 Workshop), BetaRisk (AAAI'26 Under Review)

Conceptual architecture of the Dynamic Geospatial Knowledge Engine. The engine fuses diverse, real-time data streams (left) to power a central AI core. This process transforms raw data into tailored, actionable insights for key stakeholders (right), including optimized routing for emergency responders, personalized risk warnings for the public, and data-driven policy assessments for city officials.

Trustworthy AI for Medical Image Analysis

Abstract: The clinical translation of AI for medical image analysis is impeded by fundamental challenges to its trustworthiness. Prevailing deep learning models are frequently susceptible to dataset bias, provide unreliable uncertainty estimations, and lack sufficient interpretability. These systemic limitations present significant barriers to widespread clinical adoption, as they can result in diagnostic errors and undermine clinician confidence in automated systems. This research program addresses these deficiencies through a sustained focus on designing novel algorithms for explainable AI (XAI) and robust uncertainty quantification. The core of our agenda involves developing foundational computational methods engineered for a diverse range of imaging modalities, including MRI, CT, X-ray, mammogram, and digital breast tomosynthesis. A key characteristic of our work is the architectural flexibility of these algorithms, which are designed to achieve high performance in both single-modality analysis and sophisticated multi-modal integration with auxiliary clinical data. The overarching objective of this research program is to produce ethical, equitable, and clinically robust AI systems. Through these collective efforts in algorithmic innovation, we aim to significantly improve diagnostic precision, bolster clinician trust, and ultimately, advance patient outcomes across a variety of healthcare applications.

Keywords: Calibration, Uncertainty Estimation, Explainability, Training Efficiency

Diseases: Breast Cancer, Alzheimer's Disease, Brain Tumor, Lung Nodule

Selected Papers: NN Inconsistency Performance (JACR'19), DCA for Calibration (BMVC'20), Dynamic Image for Alzheimer’s (ECCV'20 Workshop), Text-Image Pre-Training (IEEE JBHI'21), Probabilistic Calibration (IEEE BigData'24), Benchmaking Robustness (AAAI'25 Workshop)

Mitigating Bias in Class-Imbalanced Image Synthesis Models

Abstract: Image synthesis using diffusion models has made significant progress and revolutionized the content creation industry. However, text-to-image models are observed to amplify stereotypes on race and gender. For example, synthetic images for the class "lawyer" are often white males, while those for the class "cashier" are often colored female. Data-driven diffusion models require abundant images for training, while real-world images follow a long-tailed distribution with a limited number of images for rare classes. We aim to mitigate such biases for class-imbalanced diffusion models for image synthesis with two goals. Our first goal is to increase the frequency of occurrence for rare class inference. The second goal is to improve the quality of synthetic images for rare classes with a limited amount of training images while maintaining the quality for head classes.

Keywords: GenAI, Bias

Selected Papers: Negative InfoNCE (ICLR'26 Under Review), S2O Translaiton (IEEE J-STARS Under Review)

Generative Single-View 3D Reconstruction of Complex Biological Forms via Implicit Neural Representations

Abstract: This research project addresses a fundamental challenge in computer vision and computational biology: the reconstruction of high-fidelity 3D models from limited 2D visual data. Specifically, this work will develop a novel deep learning framework capable of generating a complete 3D model of an insect from a single 2D image. The current state-of-the-art in 3D modeling requires either expensive, specialized scanning equipment (e.g., micro-CT) or multi-image photogrammetry, which are labor-intensive and not scalable for rapid biodiversity assessment or digital archiving. This project will bridge this gap by creating a generative model that learns the "idea" of an insect's shape, enabling it to infer complex, occluded geometry from a single viewpoint.

More projects will be added soon...