Is Adversarial Training with Compressed Datasets Effective?