Novel Hybrid Framework Enhances Adversarial Robustness in Machine Learning Models

The increasing sophistication of cyber threats poses significant challenges to machine learning models, particularly in sensitive sectors like healthcare and finance. A new hybrid adversarially-trained deep learning framework has been developed to improve resilience against common adversarial attacks, specifically the fast gradient sign method (FGSM) and projected gradient descent (PGD). This framework integrates reinforcement learning-inspired robustness adaptation with knowledge-driven regularization, achieving 97.88% accuracy on clean data and maintaining 84.9% accuracy under FGSM and 81.75% under PGD attacks. The model outperforms traditional convolutional neural networks (CNN) and long short-term memory (LSTM) models by 6–10 percentage points in adversarial robustness. The research emphasizes the importance of balancing clean accuracy and adversarial robustness while ensuring stable convergence. The findings contribute to a deeper understanding of the trade-offs between robustness, efficiency, and generalization in machine learning applications. The dataset used for this study is publicly available on Kaggle.

Key Points: • A novel hybrid framework improves adversarial robustness in machine learning models. • The model achieves up to 97.88% accuracy on clean data and maintains over 81% under attacks. • Research highlights the importance of balancing accuracy and robustness in cybersecurity applications.