Design of a low-power constraint communication AI fusion model compression algorithm

In low-power edge communication scenarios, while AI (Artificial Intelligence) model compression can improve energy efficiency, it often introduces security vulnerabilities due to structural simplification and quantization operations, leading to decreased adversarial robustness and increased privacy risks, forming a core contradiction between “energy efficiency improvement” and “inherent security deficiency.” To address this, this paper proposes a lightweight, inherently secure communication-AI fusion model compression algorithm. The scheme integrates communication-aware structured pruning, security-aware hybrid precision quantization, inherent adversarial training, and differential privacy mechanisms. End-to-end collaborative optimization is achieved via a differentiable joint loss function for power consumption and security. This method collaboratively optimizes efficiency and security while adhering to strict power-consumption constraints. Experiments show that the proposed method consumes only 4.8 mJ per inference cycle on the STM32H743 platform, while maintaining high performance. The accuracy of member inference attacks is 52.9%, and the Normalized Mean Square Error ( NMSE ) for CSI ( Channel State Information ) reconstruction is -15.9 dB. Furthermore, the mean ΔNMSE under PGD (Projected Gradient Descent) attacks is approximately 0.93 dB. significantly outperforming existing compression baselines. This work provides a deployable compression paradigm for resource-constrained communication AI systems that balances energy efficiency, accuracy, robustness, and privacy, and has significant engineering application value.