Research on the Evaluation of Model Attack Threat Effectiveness in Complex Electromagnetic Environment

In the complex electromagnetic environment of the enemy and our two sides not only the existence of intense antagonistic interference, but also the dynamic overlapping signal will be the influence of natural noise. This paper proposes to construct a model attack threat effectiveness evaluation system in complex electromagnetic environment, which integrally considers the influence of antagonistic samples and natural noise. The robustness of the model is measured from three aspects, namely, model behavioral performance, natural noise and imperceptible perturbation, and a comprehensive visual evaluation method combining the three is proposed. Meanwhile, the evaluation system considers different confrontation attack methods and attack modes to meet the complex battlefield electromagnetic confrontation environment. The experimental results show that MobileNet is less vulnerable to one-step attacks, and Resnet is able to maintain a better ACA at low perturbations but as the perturbations increase the model becomes less able to withstand the attacks. As the PNR increases, the imperceptibility of the perturbation decreases, which leads to a decrease in the model recognition performance. The evaluation system can provide an intuitive and comprehensive evaluation of model attack threat effectiveness in complex electromagnetic environments, providing a new method for spectrum-aware models to evaluate their effectiveness against attacks.