An Adversarial Attack Approach on Financial LLMs Driven by Embedding-Similarity Optimization

Abstract

Adversarial attacks on financial sentiment analysis models are a critical area of research within NLP. We introduce a novel white-box attack method that leverages a pre-trained general-purpose language model to generate high-quality and human-imperceptible attacks. Unlike existing methods that rely on training specialized adversarial models or computationally-intensive gradient optimization routines, our approach employs carefully-designed instructions and a novel embedding-similarity function to maintain semantic integrity while producing linguistically rich adversarial samples with- out requiring ground-truth sentiment labels. Our results, obtained from attacking FinBERT and FinGPT models across three public datasets, demonstrate significant performance degradation for the two models, with MAE values reaching 0.67 and 0.48, respectively.