Fast adversarial robustness recovery from model parameter exposure through Hopfield‐type neural ordinary differential equations

Abstract Model parameter leakage transforms black‐box settings into near‐white‐box threats, enabling the insertion of highly transferable adversarial examples through exposed model internals. Defending against such attacks requires fast, robust, and minimally invasive adaptation methods. To address this challenge, we introduce a lightweight plug‐in defense module that refines corrupted latent features while preserving a frozen backbone model. Our approach leverages ordinary neural differential equations with a Hopfield‐type architecture, which is designed to purify latent features with strong stability properties. The key innovation of our model is a channel gain normalization technique that analytically binds the Lipschitz constant, thereby improving training stability and robustness. Notably, our method requires only 5% of the original training data to outperform standard adversarial training on both clean and adversarial samples, achieving 84.8% robust accuracy on CIFAR‐10 (+6.2% over adversarial fine‐tuning), while maintaining a clean accuracy of 90.4%. Furthermore, our model effectively adapts to repeated model leakages and memory‐aware attacks by supporting rapid reinitialization. Extensive experiments on the CIFAR‐10 and CIFAR‐100 datasets demonstrate the effectiveness of the proposed approach. This work represents a practical step toward fast, computationally adaptive defenses against model leakage attacks, which are becoming increasingly common in modern machine learning environments.