IDMN-SP: Improved Deep Maxout Network with Adaptive Structured Pruning for Resource-Efficient Federated Learning in IoT Big Data Analytics

Abstract

The exponential growth of Internet of Things (IoT) ecosystems generates massive heterogeneous big data, requiring efficient, privacy-preserving analytics that operate under strict edge resource constraints. Current federated learning (FL) frameworks often struggle with high computational costs, poor bandwidth, and ineffective pruning or ReLU activation bottlenecks that trigger severe accuracy losses and dead neurons. To address these challenges, this paper introduces the Improved Deep Maxout Network with Adaptive Structured Pruning (IDMN-SP), a lightweight federated learning architecture designed for resource-efficient IoT big data analytics. IDMN-SP features two core innovations: replacing traditional ReLU with a Bendable Linear Unit (BLU) function to eliminate dead neurons and enhance feature representation, and an adaptive gradient-based structured pruning mechanism employing normalized L2-norm thresholding to selectively remove redundant convolutional filters without discarding key features, paired with a targeted fine-tuning recovery phase. Evaluated on the CIFAR-10 benchmark dataset under non-IID conditions (), the proposed framework achieves an outstanding classification accuracy of 84.73%, a Precision of 84.41%, a Recall of 84.59%, and an F1-Score of 84.50%. Moreover, IDMN-SP drastically improves resource efficiency by reducing model complexity to 24.19 million FLOPs, shrinking model size to 1.94 MB, and lowering communication bandwidth to 1.94 MB per round, thereby outperforming baseline methods like FedDA by 4.56 percentage points in accuracy. These results demonstrate that IDMN-SP successfully bridges the gap between high predictive accuracy and strict resource conservation, making it an optimal solution for edge-based IoT environments.