Utility companies are facing significant challenges related to non-technical losses, which are contributing to a continuous increase in revenue loss. The application of machine learning (ML) plays a crucial role in mitigating energy theft and reducing these losses. This paper presents a hybrid method that combines a convolutional neural network (CNN) and the CatBoost algorithm. In this model, features are extracted by CNN, while CatBoost is used for classifying consumers as either legitimate or fraudulent. Additionally, this study addresses challenges such as data imbalance, missing values, etc. The ablation experiment is also conducted in order to check the individual components’ importance and performance. We also compared the model with benchmark methods. The performance assessment is carried out based on the electricity consumption (EC). The results obtained in the study show that the proposed hybrid framework delivers exceptional classification accuracy and reliability, outperforming other established and advanced machine learning models. Our findings demonstrate that the proposed system outperforms other benchmark methods, achieving an accuracy of 94.32%; we conducted an ablation experiment to understand the individual method’s contribution. We also proposed an adversarial attack to test the performance of the model, which successfully reduced the accuracy of the model up to 88.54%, compared with other attacks on benchmark methods. This system shows great potential for practical applications in detecting electricity theft in the industrial field.