Deep learning technologies have rapidly advanced, but concerns about their security and vulnerability to threats have emerged. Adversarial attacks, using carefully crafted perturbations, exploit these weaknesses, posing serious risks. This study introduces an integrated defensive preprocessing and adversarial training pipeline as a robust defense against multiple attacks on image classification models. Defensive preprocessing employs random noise, average pooling, and denoising to mitigate attack impacts, while super-resolution recovers details lost during pooling. Adversarial training generates samples under various attacks, helping the model recognize and learn from attacked images. Our results indicate that this combined approach effectively defends against multiple unknown attack types with fewer training samples. This versatile method demands minimal computational resources and doesn’t require retraining the underlying model. It can also adapt to new, unseen attacks with training on just one attack type. Experiments on standard datasets highlight the approach’s effectiveness in defending against both known and new attacks.