Improving Fairness in Deepfake Detection