Framework is not immune if over fitting problems may occur since Training data contains additional noises and training data is insufficient. To solve this case a central aggregator can call the evaluation model function based on the testing data set to solve the mentioned problems and the task publisher and the central aggregator should keep the tested data unreleased during the model update phase. In this report, we have used a differential privacy technique and imple mented this technique with existing machine learning algorithms to provide pri vacy for the trained data. In particular, we have wrapped the existing optimizer into their deferentially private counterparts using TensorFlow and MNIST data set to be able to test the differential privacy method, tuning hyper parameters introduced by deferentially private machine learning and measuring privacy guar antee provided using analysis tools included in TensorFlow Privacy. Through the advancement of data collection and retrieval technologies, the question of privacy leakage relevant to individuals is eventually revealed while revealing or exchanging data to extract valuable judgment details and expertise, it should be extensively explored on a technology basis to resolve such problems. As for future work, many techniques could be studied related to privacy preserving such as finding other possible privacy breaches based on Anonymization techniques. Also, the 33 integration of anonymous authentication mechanism that could be used in well known social media and Email applications such as Facebook,twitter, Gmail... Many proposed methods could be widely studied and helpful but in the end the privacy preserving is realized by knowledge and awareness of each individual.