Towards Trustworthy Generative AI: Privacy, Alignment, and Applications

The rapid rise of generative AI (GenAI) has enabled powerful applications in creative image generation, cross-modal synthesis, and medical decision-making. However, these capabilities come with challenges in reliability, safety, privacy, and alignment with human values. This dissertation presents methods for building GenAI systems that are trustworthy, safe, and value-aligned, with practical impact in both visual and clinical domains.

First, we address the problem of learning from negative experiences in offline reinforcement learning, especially critical in healthcare. We propose PosNegDM, a sepsis treatment model combining a transformer-based policy learner with a mortality classifier. Drawing on ideas from GANs, PosNegDM generates predictive patient states and learns to avoid high-risk outcomes, significantly reducing predicted mortality compared to existing baselines.

Second, we introduce BalancedDPO, a new framework for aligning diffusion-based generative models with multi-metric feedback, including human preferences, CLIP similarity, and aesthetic quality, using majority-vote aggregation. BalancedDPO outperforms prior methods while using 32x fewer compute resources, solving key challenges in multi-objective alignment.

Finally, we design a privacy-first image editing system, PrivateEdit, that masks identity-sensitive regions before sending data to external models, and reintegrates them post-generation using structure-aware blending. This pipeline enables secure, high-quality image editing with user control over identity exposure, and has led to a filed provisional patent and commercial interest.

Together, these contributions lay the groundwork for safer, more aligned, and privacy-preserving generative AI systems.