GAN-based classifier protection against adversarial attacks