Design of a federated learning algorithm for power big data privacy computing based on pruning technique and homomorphic encryption

This paper proposes a pruned homomorphic encryption-based collaborative federated learning algorithm to address the issues of non-uniform data distribution, overfitting caused by residual gradient features, and privacy leakage risks in power big data computation. The algorithm integrates the advantages of dynamic pruning and homomorphic encryption techniques: pruning mitigates overfitting risks and enhances efficiency by sparsifying the model (reducing the number of parameters by 30% to 50%), while homomorphic encryption ensures the security of gradient aggregation in the ciphertext domain (reducing computation cost by 60%). The specific procedure includes: initializing model parameters and generating encryption keys; designing a federated acceleration algorithm based on dynamic pruning to compress gradients and perform homomorphic encryption; updating model parameters by aggregating ciphertext gradients via weighted averaging; and finally decrypting the aggregated result to obtain the plaintext output. Experimental results demonstrate that the proposed algorithm, combined with automatic meta-pruning technology and a partial homomorphic encryption scheme, achieves a 47% improvement in computational efficiency and a 36% reduction in communication overhead under million-scale data volumes. It supports collaborative training among 50 institutions (with training time per institution for million-level data less than 8 h). By integrating encryption and pruning in a dual mechanism, the algorithm balances privacy protection and model performance, offering a secure and practical privacy-preserving computation solution for the power industry.