Quantum-Based Detection of Adversarial Attacks in NLP

This research article presents a framework for detecting adversarial attacks in Natural Language Processing (NLP) systems using Projected Quantum Kernels (PQK). The study outlines the challenges posed by adversarial attacks, which can manipulate inputs to mislead NLP models. Classical detection methods often struggle due to their computational complexity and limited generalization capabilities. The proposed approach transforms classical text embeddings into quantum states, which are then evolved to extract feature relationships before being projected back into classical space for classification using a Support Vector Machine (SVM). The framework was evaluated on a dataset of 10,896 samples from the AG News dataset, achieving an F1-score of 82% in optimal configurations. The results indicate that while quantum features can enhance detection capabilities, increasing circuit complexity can degrade performance, highlighting the importance of careful design in quantum circuits. Future work will explore comparisons with classical methods and broader testing against various adversarial attacks.