“That Is a Suspicious Reaction!”: Interpreting Logits Variation to Detect NLP Adversarial Attacks

37155

Knowledge Triples

Source Entity	Relationship	Target Entity
Adversarial Attacks Task	addressedBy	Defense Strategies Method
Defense Strategies Method	developedFor	Model Agnostic Detector Method
Model Agnostic Detector Method	analyzes	Logits OtherEntity
Model Agnostic Detector Method	evaluatedIn	Datasets Material
Model Agnostic Detector Method	improves	State of the Art Performance Metric