Continual Learning with Adversarial Training to Enhance Robustness of Image Recognition Models
Ting-Chun Chou, Jhih-Yuan Huang, Wei‐Po Lee · 2022
In recent years, deep learning has been widely adopted in many image recognition tasks with great success. However, this method is vulnerable against well-designed attacks, causing serious safety and security problems in life- critical applications. To overcome the issue of continuously evolving attacks, in this work, we develop a new defensive approach that integrates continual learning and adversarial training to improve both corruption robustness and structure compactness of the defensive model. Our approach adopts the structure of progressive neural model to establish a robust model over time. It includes adaptive phases of model growing and pruning in continual learning, and performs adversarial training iteratively during the learning process. To evaluate the performance of the proposed approach, we conduct a series of experiments to compare our approach with others in defending the current well-known adversarial attacks. The results that our model can obtain best performance and provide an effective approach in cybersecurity.