In machine learning, differential privacy is applied by integrating privacy- preserving techniques into the training and inference stages of models. One common method is differentially private stochastic gradient descent (DPSGD), where noise is added to the gradient updates during training. This ensures that the final model doesn't retain or reveal sensitive information about individual data points. The approach balances the need for model accuracy with privacy considerations. By carefully controlling the amount of noise and the frequency of data access, machine learning models can be trained on sensitive datasets while providing strong privacy guarantees.