Adversarial attacks aim to deceive deep learning models by providing misleading or altered input data, exposing weaknesses in AI systems.
In Natural Language Processing (NLP), adversarial attacks involve making subtle changes to text that confuse AI models while appearing normal to humans.
These attacks can occur at the character level, word level, or sentence level, and can be combined for higher effectiveness.
Generating imperceptible adversarial attacks in NLP is challenging due to the noticeable nature of text alterations.