An interpretable deep learning model-based adversarial defense method

Abstract

This invention belongs to the field of explainable techniques for deep neural network models, specifically an interpretable deep learning model-based adversarial defense method. The method involves establishing an adversarial model, utilizing the model for adversarial training to enhance model robustness. It's remarkably straightforward to use, requiring only the generation of adversarial samples through existing methods without altering the structure or training methods of the current model. From a quantitative perspective, it compares the differences between two networks through accuracy testing. This invention empowers the post-training network with improved robustness, mitigating the impact of adversarial samples on the classification and prediction of deep neural network models. Its existence brings greater certainty to the technology itself and its practical applications, allowing the obtained model to render visual explanations that closely approximate both real samples and adversarial samples.