Contextual Embedding Inversion Attack (CEIA): A Novel Adversarial Method Targeting BERT Models

This article describes CEIA, a new, targeted adversarial attack on BERT, name as Contextual Embedding Inversion Attack. CEIA uses BERT for reversing the general meaning of the crucial phrases based on the context’s proximity while maintaining syntactical integrity. This method is entirely new in the permutation of phrases such that BERT’s contextual embeddings are used to generate contrary phrases with an especial concentration on noun chunks. Due to transforming most of the words’ meaning while preserving the grammatical structure, this method represents an essential challenge for BERT-based models addressing the selected texts.