Paper accepted for ACL 2022: “That Is a Suspicious Reaction!”: Interpreting Logits Variations to Detect NLP Adversarial Attacks.