Publications by authors named "Angelo Sotgiu"
Domain Knowledge Alleviates Adversarial Attacks in Multi-Label Classifiers.
IEEE Trans Pattern Anal Mach Intell
December 2022
Article Synopsis
- The paper examines adversarial attacks and defenses in multi-label classification, highlighting how domain knowledge can help identify incoherent predictions caused by these attacks.
- By integrating first-order logic constraints into a semi-supervised learning framework, the authors demonstrate that classifiers can reject samples that don't align with the established domain knowledge.
- Their findings reveal that even without prior knowledge of specific attacks, domain constraints can effectively detect adversarial examples, suggesting a path toward more resilient multi-label classifiers.