Abstract
Deep learning (DL)-based specific emitter identification (SEI) efficiently enables automatic device recognition in consumer electronics for intelligent content retrieval and personalized services. However, adversarial attacks can manipulate wireless signal features to deceive DL-based models of SEI, compromising device authentication and system security. Therefore, this paper proposes a meta-learning-based self-adaptive adversarial defense (MSAAD) for SEI, which can dynamically predict adversarial perturbations and sample defense weights. Specifically, the predicted perturbations guide adversarial training to generate adversarial attacks with different intensities for different samples, enabling fine-grained robustness against varying attack intensities. Meanwhile, the predicted defense weights adaptively balance the loss contributions from both clean and adversarial examples. Adaptive parameter adjustment results in a more targeted defense. Experiment results indicate that the proposed adaptive defense approach achieves a better adaptability to various adversarial examples. The MSAAD yields up to a 64% performance gain over the Baseline, and surpasses several state-of-the-art methods by as much as 58% under different signal-to-noise ratios and different attacks, substantially strengthening the adversarial robustness of the SEI model, to ensure reliable device authentication for secure intelligent content retrieval and personalized services in consumer electronics.
Funding Information
  • National Natural Science Foundation of China (62331010)
  • Fundamental Research Funds for the Central Universities through Harbin Engineering University (HEU) Project (3072024XX0806)
  • Fundamental Research Funds for the Central Universities through Harbin Engineering University (HEU) Project (3072025ZX0802)

This publication has 50 references indexed in Scilit: