Nowadays, more and more machine learning applications, such as medical diagnosis, online fraud detection, email spam filtering, etc., services are provided by cloud computing. The cloud service provider collects the data from the various owners to train or classify the machine learning system in the cloud environment. However, multiple data owners may not entirely rely on the cloud platform that a third party engages. Therefore, data security and privacy problems are among the critical hindrances to using machine learning tools, particularly with multiple data owners. In addition, unauthorized entities can detect the statistical input data and infer the machine learning model parameters. Therefore, a privacy-preserving model is proposed, which protects the privacy of the data without compromising machine learning efficiency. In order to protect the data of data owners, the epsilon-differential privacy is used, and fog nodes are used to address the problem of the lower bandwidth and latency in this proposed scheme. The noise is produced by the epsilon-differential mechanism, which is then added to the data. Moreover, the noise is injected at the data owner site to protect the owners data. Fog nodes collect the noise-added data from the data owners, then shift it to the cloud platform for storage, computation, and performing the classification tasks purposes.