Improving robustness of jet tagging algorithms with adversarial training: exploring the loss surface