MEAT: Median Ensemble Adversarial Training for Improved Robustness and Generalization

MEAT (Median Ensemble Adversarial Training) eliminates outliers in weight space by searching for the median of historical model weights, addressing the robust overfitting problem that still exists in self-ensemble adversarial training methods such as weight averaging (WA). The method achieves the best robustness under AutoAttack and effectively mitigates robust overfitting. Most defense methods, when combined with MEAT, can also improve robust generalization ability and robustness.