In particular, machine learning and data analytics applications are moving an increasing amount of data storage and computation from the local to the cloud due to the rapid development of cloud computing. However, because the cloud servers are managed by a third party, consumers cannot totally rely on them. As a result, it becomes difficult to implement privacy-preserving deep learning over cloud data from several data sources. The Differential Privacy based Cloud Security Model (DP-DSM), which secures both cloud data sets and the data sets of multiple providers, is thus proposed in this study as a novel paradigm. The model makes use of a differential privacy algorithm to protect the privacy requirements of various providers. Additionally, the model redesigns the training procedure and requires the data owner to include a randomization layer before transmitting the data in order to counteract an unreliable deep learning model. The Convolutional Neural Network (CNN) architecture is divided into three layers: the convolutional unit, the randomization unit, and the fully connected unit. In addition, the proposed approach is assessed for time consumption and attack detection accuracy, making it more useful for IoT-driven cloud environments than previous methods.