Adaptive Feature Distillation-Based Continuous Authentication Against RF Fingerprint Drift for Power Equipment

RF fingerprints of power equipment drift over time due to environmental changes and device aging, which progressively degrades the performance of authentication systems built on fixed models. Existing incremental learning methods tend to either forget historical devices or fail to adapt to new distributions when dealing with such drift. We propose an incremental update strategy tailored to this scenario and evaluate it on a self-constructed 64-dimensional simulated drift dataset. At each update, gradient importance per feature channel is computed from the current batch, and three feature-level distillation terms (channel MSE, covariance alignment, and spatial attention) keep the new model’s representations close to the old one. The distillation strength decays exponentially with update steps, enabling strong preservation of old knowledge early and more flexible adaptation later. A small memory buffer mixes old samples into each training batch to further reinforce historical recognition. On the synthetic dataset, our method achieves higher final historical accuracy than static, fine-tuning, EWC, and LwF baselines. Ablation studies confirm that the multi-level distillation, dynamic decay, and momentum-smoothed channel weights each contribute positively. These preliminary simulation-based results indicate that the method shows promise in alleviating forgetting caused by fingerprint drift while maintaining adaptability to new fingerprints within the synthetic evaluation framework.