In this work, we implement Adversarial Machine Learning (ML) for bias mitigation in models with one and two sensitive features, using Demographic Parity and Equalized Odds as fairness metrics. To address the challenges of multi-sensitive-feature settings, we introduce updated formulations for these metrics and modify loss functions to improve both fairness and predictive performance. Additionally, we incorporate a previously proposed heuristic approach, stacking, to enhance bias mitigation. Our results show that for models with a single sensitive feature, stacking improves accuracy while maintaining comparable discrimination levels. In contrast, for models with two sensitive features, stacking significantly reduces discrimination without compromising accuracy. To ensure real-world applicability, we evaluate our models on imbalanced datasets, reflecting disparities commonly found in domains such as hiring and loan approvals. These findings highlight the effectiveness of integrating adversarial ML with refined fairness metrics and loss functions, demonstrating a scalable and practical solution for mitigating bias in complex real-world scenarios.