Abstract:Security threats such as adversarial and supply chain attacks are increasingly prominent in deep learning-based intelligent systems. Traditional systems typically adopt a single model with static and deterministic defense mechanisms, making them susceptible to single points of failure and lacking in security resilience. To address this issue, a method termed N-Model is proposed, which leverages the dynamic combination of multiple deep learning models to promote diversity and randomness. The dynamic switching among models increases the uncertainty of attack targets and methods. A multi-model voting mechanism is integrated to further strengthen security resilience. Theoretical analysis indicates that the proposed N-Model achieves higher expected accuracy than single-model approaches under attack scenarios. Experimental evaluations on the CIFAR-10 dataset demonstrate that the N-Model effectively resists various adversarial attacks, exhibiting a lower attack success rate and superior overall security performance.