I aim to create socially responsible machine learning (ML) models, i.e., ML models that can protect the privacy of their training data and minimize the chance of being misused. In particular, to ensure privacy protection, those models must be resistant to various privacy attacks. To prevent harmful social impacts from the misuse of ML models, their generalization ability on unintentional data domains and tasks must be reduced. To achieve these goals, I have been developing novel ML models and tools with a diverse set of techniques in Optimization Theory, Information Theory, and Cryptography. Greater ability comes with greater responsibility, and this applies equally to ML technology. ML models with social responsibility can protect data privacy and prevent serious consequences caused by the harmful use of models, such as teenagers obtaining violent information from recommendation systems or criminals using ML models to engage in criminal activities. In particular, my current research lies in Model Protection, Federated Learning, and Crypto-ML.