AdaFed is a fair federated learning (FL) method that adaptively tunes a common descent direction for the server to ensure fairness among clients. The goal of AdaFed is to find an updating direction for the server such that (i) all clients' loss functions decrease, and (ii) clients with larger loss values decrease more rapidly. AdaFed adaptively adjusts this common direction based on local gradients and loss functions. The method is validated on a suite of federated datasets, demonstrating that AdaFed outperforms state-of-the-art fair FL methods in terms of fairness while maintaining similar prediction accuracy. In FL, the server aggregates local gradients from clients to update the global model. However, this common direction may not be descent for all clients, leading to unfair performance. To address this, AdaFed treats FL as a multi-objective minimization problem, finding a common descent direction that is descent for all clients and ensures that loss functions with larger values decrease more rapidly. This is achieved by using directional derivatives and adaptive scaling of gradients. AdaFed's methodology involves two phases: orthogonalization of gradients and finding the optimal scaling factor. In the first phase, gradients are orthogonalized to ensure that the common direction is descent for all clients. In the second phase, the optimal scaling factor is determined to ensure that the common direction is more inclined toward clients with larger loss functions, leading to higher rates of decrease for those clients. Theoretical analysis shows that AdaFed converges to a Pareto-stationary solution under certain conditions, ensuring fairness and convergence. Experimental results on seven datasets, including CIFAR-10, CIFAR-100, FEMNIST, and Shakespeare, demonstrate that AdaFed achieves higher fairness and comparable accuracy compared to existing fair FL methods. The method is also shown to be effective in scenarios with high non-iidness, where client data distributions differ significantly. The results indicate that AdaFed provides a more fair and efficient solution for federated learning tasks.AdaFed is a fair federated learning (FL) method that adaptively tunes a common descent direction for the server to ensure fairness among clients. The goal of AdaFed is to find an updating direction for the server such that (i) all clients' loss functions decrease, and (ii) clients with larger loss values decrease more rapidly. AdaFed adaptively adjusts this common direction based on local gradients and loss functions. The method is validated on a suite of federated datasets, demonstrating that AdaFed outperforms state-of-the-art fair FL methods in terms of fairness while maintaining similar prediction accuracy. In FL, the server aggregates local gradients from clients to update the global model. However, this common direction may not be descent for all clients, leading to unfair performance. To address this, AdaFed treats FL as a multi-objective minimization problem, finding a common descent direction that is descent for all clients and ensures that loss functions with larger values decrease more rapidly. This is achieved by using directional derivatives and adaptive scaling of gradients. AdaFed's methodology involves two phases: orthogonalization of gradients and finding the optimal scaling factor. In the first phase, gradients are orthogonalized to ensure that the common direction is descent for all clients. In the second phase, the optimal scaling factor is determined to ensure that the common direction is more inclined toward clients with larger loss functions, leading to higher rates of decrease for those clients. Theoretical analysis shows that AdaFed converges to a Pareto-stationary solution under certain conditions, ensuring fairness and convergence. Experimental results on seven datasets, including CIFAR-10, CIFAR-100, FEMNIST, and Shakespeare, demonstrate that AdaFed achieves higher fairness and comparable accuracy compared to existing fair FL methods. The method is also shown to be effective in scenarios with high non-iidness, where client data distributions differ significantly. The results indicate that AdaFed provides a more fair and efficient solution for federated learning tasks.