Abstract: Aiming at the problem that the intrusion detection system based on deep neural network shows insufficient robustness under diverse adversarial attacks and the traditional fusion method is difficult to take into account the characteristics of different attack scenarios,this paper proposes a dynamic adversarial sample detection framework based on multi-modal uncertainty fusion.Firstly,a training progress aware dynamic loss scheduling mechanism is designed to realize the balance optimization between feature learning and defense robustness by linearly scheduling the weights of normal samples and adversarial samples.Secondly,a collaborative fusion mechanism of uncertainty and attention is proposed.Monte Carlo dropout is used to construct the prediction distribution and quantify the uncertainty of the model.Then,the context-sensitive attention module is combined to adaptively adjust the contribution weights of each sub-model to realize the dynamic weighted integration of multimodal features,so as to effectively suppress the anti-noise and highlight the key features.Experimental results show that the proposed framework achieves more than 96% detection accuracy and nearly 100% AUC under normal samples and multiple attack samples(FGSM,PGD,BIM),which is significantly better than AT-DNN and Roshan and other representative adversarial defense methods,fully veri-fying the effectiveness of the proposed framework in improving robustness and generalization ability.

Key words: Adversarial detection, Dynamic loss scheduling, MC dropout, Multi-head attention, Multimodal uncertainty fusion