ZDDR: A Zero-Shot Defender for Adversarial Samples Detection and Restoration

• Published in: IEEE access Vol. 12; pp. 39081 - 39094
• Main Authors: Chen, Musheng, He, Guowei, Wu, Junhua
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 2024; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Adversarial defense; Adversarial machine learning; Computational modeling; Data models; Detection algorithms; large language model; Large language models; model security; Natural language processing; Performance standards; Perturbation methods; prompt engineering; Restoration; Robustness; Semantics; Sentences; Training

Natural language processing (NLP) models find extensive applications but face vulnerabilities against adversarial inputs. Traditional defenses lean heavily on supervised detection techniques, which makes them vulnerable to issues arising from training data quality, inherent biases, noise, or adversa...