Ensuring Fairness Beyond the Training Data