Multifeature Fusion-Based Closed-Box Attack Detection Method for Automatic Modulation Classification

The application of deep learning (DL) technologies in automatic modulation classification (AMC) faces significant challenges from adversarial attacks. Existing attack methods often rely on idealized assumptions and are inadequate for complex real-world scenarios. To address this issue, we propose an innovative multifeature fusion-based closed-box attack detection method (MFCA). This method enhances the effectiveness and quality of adversarial examples by utilizing a multimodel fusion strategy to extract diverse features. MFCA consists of three key steps: 1) we obtain an initial dataset from the target model and use a multifeature fusion model to approximate its decision boundary; 2) we enhance the model’s stability by applying data augmentation and incremental learning techniques through multiple rounds of training; and 3) we generate adversarial examples using the optimized fusion model. By constructing this comprehensive attack framework, we demonstrate how integrating diverse model characteristics enhances adversarial attack performance and evaluate the robustness of the MFCA method against defense strategies. Experimental results show that adversarial examples generated by MFCA exhibit significantly higher transferability across different perturbation levels compared to those from single substitute models and other attack methods, with superior attack efficacy. This research provides an effective solution for closed-box adversarial attacks and validates the advantages of multifeature fusion models in adversarial example generation.