My research has been primarily focused on the development of robust and efficient Bayesian Neural Networks (BNNs) and their hardware realizations, aiming to address the growing challenges in adversarial robustness, model uncertainty quantification, and energy-efficient AI inference. Central to my work is the integration of Bayesian inference with deep learning models, such as WideResNet and VGG architectures, in order to enhance their generalization under distributional shifts and adversarial perturbations.One of the major contributions of this research lies in the design and implementation of a selective Bayesianization framework, which enables the dynamic integration of stochastic components—such as Bayesian convolutional layers—into pre-trained deterministic networks. This framework significantly improves model robustness while controlling computational cost, and it provides the flexibility to switch between deterministic and stochastic modes during inference, thereby supporting energy-aware decision-making.In parallel, I have developed adversarial training strategies tailored for Bayesian models. These methods include sampling-based perturbation alignment and uncertainty-guided adversarial filtering, which leverage the posterior variance of model predictions to enhance resilience against gradient-based white-box attacks, such as FGSM, PGD, and CW. Empirical evaluations on benchmark datasets such as CIFAR-10 and CIFAR-100 demonstrate that the proposed Bayesian models outperform their deterministic counterparts under a wide range of threat models.LSIとシステムのワークショップ2025  学会发表