The paper addresses the issue of fairness generalization in deepfake detection, where existing methods often fail to maintain fairness across different demographic groups and domains. The authors propose a novel framework that combines disentanglement learning, fair learning, and optimization to improve fairness generalization. Disentanglement learning extracts demographic and domain-agnostic forgery features, which are then fused to promote fair learning. The framework uses a bi-level fairness loss to enhance fairness both within and across subgroups. Additionally, a loss flattening technique is employed to ensure the model can escape suboptimal solutions, thereby improving fairness generalization. Extensive experiments on various deepfake datasets demonstrate the effectiveness of the proposed method, showing superior performance in preserving fairness during cross-domain deepfake detection compared to state-of-the-art approaches. The code for the method is available at <https://github.com/Purdue-M2/Fairness-Generalization>.The paper addresses the issue of fairness generalization in deepfake detection, where existing methods often fail to maintain fairness across different demographic groups and domains. The authors propose a novel framework that combines disentanglement learning, fair learning, and optimization to improve fairness generalization. Disentanglement learning extracts demographic and domain-agnostic forgery features, which are then fused to promote fair learning. The framework uses a bi-level fairness loss to enhance fairness both within and across subgroups. Additionally, a loss flattening technique is employed to ensure the model can escape suboptimal solutions, thereby improving fairness generalization. Extensive experiments on various deepfake datasets demonstrate the effectiveness of the proposed method, showing superior performance in preserving fairness during cross-domain deepfake detection compared to state-of-the-art approaches. The code for the method is available at <https://github.com/Purdue-M2/Fairness-Generalization>.