Cluster-Aware Structured Pruning for Robust and Sustainable Edge Intelligence

Edge intelligence for sustainable 6G must operate under strict constraints on latency, energy, and computational availability, motivating structured pruning for efficient on-device inference. However, at high pruning ratios, current global magnitude-based channel pruning solutions remove entire feature groups, leading to severe accuracy collapse. To cope with this, we present PKPrune (PCA-KMeans Pruning), a data-free cluster-aware structured pruning method that groups convolutional filters in weight space and prunes within each group while keeping at least one representative per cluster. Ablation study reveals that clustering granularity plays a dominant role in preserving robustness under extreme pruning, while the PCA variance retention ratio has a comparatively minor impact on post-pruning accuracy. To assess practical deployment feasibility, we further study the impact of the fine-tuning budgets, showing catastrophic accuracy collapse without adaptation and rapid recovery with lightweight retraining, which enables explicit accuracy and energy trade-offs under constrained edge resources. Finally, we evaluate PKPrune on CIFAR-10 using ResNet-18 with pruning ratios up to 90%. At 90% pruning, PKPrune achieves 72.94% Top-1 accuracy after lightweight fine-tuning, compared to 63.72% with conventional L2-norm structured pruning, while exhibiting competitive efficiency behavior relative to a dependency-aware structured pruning baseline in our measurement setting.