An Adversarial Sample Generation Method Based on the Balance Between Aggression and Concealment

Aggression and concealment in adversarial sample generation present a fundamental trade-off. To achieve an optimal balance, this paper introduces input transformation and gradient stabilization strategies into the pre-convergence adversarial sample generation framework, proposing the Stabilized Momentum with Input Transformation and Preconvergence FGSM (SPMI-FGSM) method. A measurement model for aggression and concealment is constructed, and the non-dominated sorting whale optimization algorithm is employed to solve the model, generating a set of adversarial samples that satisfy the dominance relationship. The optimal equilibrium point parameters are determined using a Constant False Alarm Rate (CFAR) detector to quantify critical conditions. Experimental results demonstrate that the proposed method reduces model recognition accuracy to 2.03 % (attack condition) while maintaining a structural similarity index (SSIM) of at least 0.519 (concealment condition), with a corresponding perturbation factor of 0.09, achieving an optimal balance between attack and concealment.