In an attempt to cope with the increasing number of trust-related attacks, a system that analyzes the whole social transaction in real-time becomes a necessity. Traditional systems cannot analyze transactions in real-time and most of them use machine learning approaches, which are not suitable for the real-time processing of social transactions in the big data environment. Therefore, in this paper, we propose a novel deep learning detection system based on Apache Spark that is capable of handling huge transactions and streaming batches. Our model is made up of two main phases: the first phase builds a supervised deep learning model to classify transactions (either benign transactions or malicious transactions). The second phase aims to analyze transaction streams using spark streaming, which transforms the model to batches of data in order to make predictions in real-time. To verify the effectiveness of the proposed system, we implement this system and we perform several comparison e xperiments. The obtained results show that our approach has achieved more satisfactory efficiency and accuracy, compared to other works in the literature. Thus, it is very suitable for real-time detection of malicious transactions with large capacity and high speed.