Artificial intelligence (AI) systems play significant roles in decision making processes, yet concerns persist about potential biases that can lead to unfair outcomes. These biases can arise from two main sources: imbalanced training data distribution and correlations between sensitive attributes (such as race and gender) and the target variable. Conventional model training methods penalize performance for underrepresented groups, resulting in biased outcomes. Further, they may capture features related to sensitive attributes during training, thus exacerbating bias. Biased predictions can have detrimental consequences. To address these concerns, in this research, we propose a novel method for bias mitigation. The proposed method aims to learn fair latent representations by emphasizing task relevant features while suppressing those linked to sensitive attributes. Additionally, we employ an adaptive reweighing technique to balance target class labels during training. The proposed method is evaluated on prominent benchmark datasets and compared with existing algorithms to demonstrate its effectiveness toward bias mitigation.