How to Write an Abstract

[Background] Deep learning models have achieved remarkable success in image classification, but their vulnerability to adversarial examples—imperceptible perturbations that cause misclassification—remains a critical security concern for deployment in safety-critical applications.

[Objective] This paper presents a novel defense mechanism that improves model robustness against adversarial attacks without sacrificing accuracy on clean inputs.

[Methods] We introduce adversarial training with curriculum learning (ATCL), which progressively exposes models to increasingly sophisticated attacks during training. Our method uses a difficulty scheduler that automatically adjusts attack strength based on model performance, enabling stable training across 50,000 iterations on standard hardware.

[Results] Experiments on CIFAR-10 and ImageNet demonstrate that ATCL improves adversarial accuracy by 23% and 18% respectively compared to standard adversarial training, while maintaining 99.1% of baseline clean accuracy. The learned defenses generalize to unseen attack types, including attacks developed after training. Ablation studies confirm that curriculum scheduling is essential for these gains.

[Conclusion] These results suggest that curriculum-based adversarial training offers a practical path toward deploying robust deep learning models in security-critical applications such as autonomous vehicles and medical diagnosis.