PID: A Parameter-Efficient Isolation Domain-Incremental Learning Framework for Signal Modulation Classification.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Communications Engineering

Background:

• Deep neural networks (DNNs) show promise for signal modulation classification (SMC).
• Traditional SMC methods struggle with continuous data distribution shifts and catastrophic forgetting.
• Real-world applications like cognitive radio and cyber defense require adaptive SMC.

Purpose of the Study:

• To propose the first domain-incremental learning (DIL) paradigm for signal modulation classification (SMC).
• To develop a parameter-efficient method (PID) that enables rapid adaptation to new scenarios while retaining performance on previous ones.

Main Methods:

• Introduced a parameter space decomposition-based classifier (PSD) to separate model parameters into bases and coefficients.
• Froze bases and fine-tuned low-dimensional coefficients to mitigate catastrophic forgetting.
• Designed a scene-aware domain controller (SDC) to select domain-specific coefficients for each sample.

Main Results:

• The proposed parameter-efficient isolation (PID) method significantly reduces catastrophic forgetting in SMC.
• PID enables SMC models to adapt quickly to new communication scenarios.
• Achieved state-of-the-art (SOTA) overall performance in extensive experiments.

Conclusions:

• The novel DIL paradigm and PID method offer an effective solution for adaptive SMC in dynamic environments.
• PID successfully balances adaptation to new domains with retention of knowledge from previous domains.
• The approach demonstrates superior performance and robustness for real-world communication signal classification.