Adversarial attacks involve the provision of adversarial examples of the model, which are inputs designed by the attacker to deliberately deceive the model to produce erroneous output.