Skip to content
Open access

AGH-MGC: A Unified Compression Training Framework for Interference Recognition Models in Satellite Communications

2026 · IEEE Transactions on Cognitive Communications and Networking · Vol 12, pp. 13040-13051 · 0 citations · 55 references

Abstract

The growing complexity of satellite communication systems demands lightweight deep models capable of efficient interference recognition under strict resource constraints. However, current compression training approaches face intrinsic challenges, including the non-differentiability of hard-mask pruning, the lack of adaptive sparsity control, and gradient interference between pruning and quantization. To overcome these limitations, this paper introduces a unified compression training framework with Adaptive Gradient Harmonization and Mask-Aware Gradient Compensation (AGH-MGC), which seamlessly integrates differentiable pruning, adaptive sparsity learning, and cooperative quantization. The framework leverages Gumbel–Softmax for differentiable structure selection, employs a swarm intelligence–based strategy to dynamically regulate sparsity, and adopts a coarse-to-fine quantization scheme to optimize error allocation. Moreover, an adaptive gradient harmonization mechanism is designed to mitigate gradient conflicts and stabilize joint optimization. Experimental results demonstrate that AGH-MGC achieves an overall compression ratio of up to 96.10% across four interference recognition models, while maintaining competitive accuracy and robust optimization stability. These findings highlight the framework’s potential as a general and effective solution for structured model compression in satellite communication systems.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.