Research & Projects
Underconfidence Adversarial Training
Addressed an overlooked vulnerability in adversarially trained models: attacks that reduce model confidence without changing predictions. Decreased confidence can cause unnecessary interventions, delayed diagnoses, and erosion of trust — especially in high-stakes domains like medical imaging and autonomous systems.
Key contributions:
- Introduced two underconfidence attacks: class-pair ambiguity attacks and ConfSmooth, which spreads uncertainty across all classes.
- Developed Underconfidence Adversarial Training (UAT), integrating these attacks into standard adversarial training.
- UAT matches or beats traditional adversarial training while using half the gradient steps.
- Evaluated across 6 architectures (CNNs and Vision Transformers) and 7 datasets including MNIST, CIFAR, ImageNet, MSTAR, and medical imaging.
Model-Based Robust Training
Investigated how deep models handle natural image corruptions like snow and rain in safety-critical settings, using learned corruption models to generate realistic distortions while balancing robustness, calibration, and compute.
Key contributions:
- Comparative framework evaluating model-based training against Vanilla, Adversarial Training, and AugMix across corruption severities.
- Hybrid strategies combining random coverage with adversarial refinement in nuisance space.
- Multi-dimensional analysis of accuracy, calibration, and computational cost — beyond accuracy-only metrics.
- Showed model-based augmentation matches adversarial robustness at significantly lower cost.
Robust Training for Medical Imaging Classification
Developed robust training strategies to strengthen diagnostic neural networks against adversarial attacks and distribution shift while preserving diagnostic accuracy and clinician trust.
Key contributions:
- Proposed Robust Training with Data Augmentation (RTDA) tailored to medical imaging.
- Benchmarked across 6 baselines spanning isolated and combined adversarial training and augmentation.
- Validated across mammograms, X-rays, and ultrasound.
- Evaluated against both adversarial perturbations and natural distribution variations while preserving clean accuracy.
Research Interests
- Agentic AI — autonomous, tool-using AI systems and their reliability.
- World Models — learned models of environment dynamics for perception and planning.
- AI Assurance — trustworthy, reliable AI under adversarial and distribution stress, including adversarial machine learning, robustness to natural variation, uncertainty quantification, out-of-distribution generalization, and self-supervised representation learning.