Data leakage in machine learning, especially train–test contamination, is a serious issue that undermines the reliability of the model and reduces the applicability in the real world. When information from the test set inadvertently influences the training process, models may achieve artificially high-performance during evaluation but fail to generalize effectively in practice. To address this problem, we propose a structured framework based on adversarial validation for detecting and mitigating data leakage. The approach involves combining the training and test datasets, labeling them accordingly, and training a binary classifier to distinguish between the two. If the classifier demonstrates predictive performance substantially above chance level (AUC > 0.5), it indicates distributional discrepancies suggestive of leakage. Feature importance analysis is then applied to identify which variables are responsible for this separation. Once detected, corrective strategies such as eliminating problematic features, resampling data, or applying robust partitioning methods like time-based or group-based splits are employed to minimize contamination. The adversarial classifier is retrained after mitigation to validate the effectiveness of these interventions. This framework offers a practical and systematic solution for strengthening model robustness, improving predictive accuracy, and ensuring trustworthy deployment. By prioritizing leakage prevention, it empowers practitioners to develop interpretable and reliable machine learning systems across diverse application domains.