Lightweight Specific Emitter Identification via Joint Compression Based on Reinforcement Learning

The proliferation of spoofed signals in IoT systems, particularly in air traffic management, poses a significant threat to aviation safety, especially in resource-limited environments where computational efficiency and secure sensor networks are crucial. Current deep learning-based specific emitter identification (SEI) methods often perform pruning and depthwise separable convolution (DSC) in separate stages, neglecting their interdependencies, which results in suboptimal compression and compromised model performance. To address this, we propose a reinforcement learning-based Automated Joint Compression Network (AJCN) for efficient aircraft authentication using ADS-B signals. Our approach unifies channel pruning and depthwise separable convolution, guided by a proximal policy optimization algorithm to adaptively identify high-quality compression strategies that jointly enhance efficiency and preserve accuracy. Experiments on the public ADS-B dataset, our pruned network cuts parameters from 0.38 M to 0.03 M and FLOPs from 24.63 M to 1.15 M with only a 0.68 % accuracy loss. On the USRP dataset, it compresses even further—from 0.37 M to 0.006 M parameters and from 24.62 M to 0.63 M FLOPs while the accuracy drops by just 0.17%. The code is available at https://github.com/wdenxw/AJCN.git.