Generalised Eigenvalue Geometry of Semantic Adversarial Attacks
Researchers have developed a new theoretical framework to understand semantic adversarial attacks on machine learning models, particularly in financial sentiment classification. The work introduces a continuous local model that captures the interaction between a paraphrase and a target model, showing that the worst-case displacement is determined by the largest generalized eigenvalue of a matrix pencil derived from the models' Jacobians. This framework provides an attackability index and supports theoretical guarantees for detecting such attacks, connecting discrete search methods with continuous theory. AI
IMPACT Provides a theoretical foundation for understanding and mitigating semantic adversarial attacks on AI models.