<p>Electricity theft is a significant issue that causes substantial financial losses for utility providers and poses safety risks to the public. Detecting such theft accurately and efficiently remains a critical challenge, especially in areas with inadequate infrastructure. This research proposes a machine learning-based approach to address the problem of electricity theft, utilizing the adaptive synthetic sampling (ADASYN) technique to balance imbalanced datasets, particularly for theft detection. The study evaluates several machine learning algorithms integrated with ADASYN, including Naive Bayes, logistic regression, support vector machine, AdaBoost, decision tree, and random forest. The models were tested on publicly available datasets, yielding accuracy rates ranging from 52 to 95%. Among these, random forest demonstrated the best performance, achieving an accuracy of 95%, along with precision, recall, and F1 scores of 95% for both the “Not Theft” and “Theft” classes. The random forest model outperformed benchmarks from previous studies, showcasing its effectiveness in distinguishing between theft and non-theft instances. These results highlight the potential of machine learning models, particularly when augmented with data balancing techniques like ADASYN, to enhance the accuracy and reliability of electricity theft detection systems, thereby reducing financial losses and improving public safety.</p>