AFD-IL: A Long-Term Incremental Learning Approach With Adaptive Feature Distillation for Specific Emitter Identification

Specific emitter identification (SEI) plays a critical role for device authentication in the Industrial Internet of Things (IIoT). In real-world industrial scenarios, the continuous emergence of new devices poses significant challenges to traditional SEI systems, which often suffer from catastrophic forgetting due to their inability to incrementally adapt. Class incremental learning (CIL) offers a promising direction, but existing CIL-based SEI approaches are hampered by accumulated gradient conflicts across incremental stages, which severely degrade knowledge during long-term incremental updates. To tackle these challenges, we propose a adaptive feature distillation-enabled incremental learning (AFD-IL) method. Specifically, AFD-IL analyzes the correlation between feature distribution shifts and loss variations, and introduces an importance-aware feature distillation strategy to mitigate distributional shifts in the representations of old devices. Furthermore, we introduce a local similarity-based classifier to decouple decision boundaries between old and new devices, thereby alleviating the issue of sample imbalance. To the best of our knowledge, this is the first work to address SEI in a long-term CIL setting. Experiments conducted on the real-world ADS-B-100 dataset demonstrate that the proposed method achieves a final accuracy of 87.75% and a final forgetting rate of 9.91% after 16 incremental updates from an initial set of 20 classes. Compared to state-of-the-art CIL methods, it achieves 24.65% higher accuracy and 21.78% lower forgetting rate. These results confirm the effectiveness of the proposed method in enhancing long-term CIL performance.