Telecom fraud is a big issue that impacts operators and telecom organizations all around the world. SIMBox/Bypass fraud is one of the most common types of telecom fraud, and it involves the use of voice over IP (VoIP) technologies to avoid access charges and profit on international calls. Due to the dynamic nature of this fraud, it can quickly defeat the Test Call Generators (TCG) and FMS (Fraud Management System). In addition to TCG, a near-real-time machine learning approach has been designed to detect and classify SIMbox/Bypass fraud, which requires up-to-date data. A problem and supervised machine learning algorithms show limitations in capturing fraud’s dynamic nature. To detect and classify SIMBox/Bypass fraud, deep learning/deep neural network classifier algorithms were used in this study, namely MLP (multilayer perceptron), RNN (recurrent neural Network), and LSTM (long short-term memory), with the two validation techniques 10-fold cross-validation and separate train test. After collecting Call Detail Record (CDR) data from Ethio-Telecom, Ethiopia, relevant features were chosen, and preparation operations such as data cleaning, integrating, and aggregating were completed. The experimental results demonstrated that the MLP classifier with a separate test data validation technique obtains a higher classification accuracy of 99.17%; LSTM is ranked second with an overall performance accuracy of 98.90%; and RNN is ranked third with an accuracy of 98.10%. Detecting and classifying SIMBox/Bypass fraud has numerous advantages for telecom companies, including raising awareness to mitigate risk, preventing financial loss, increasing revenue assurance, improving operational efficiency, improving customer satisfaction, protecting brand reputation, ensuring regulatory compliance, and establishing industry leadership.