Noise-Robust Model Compression via Magnitude Weighted Hessian Pruning and ASAM Training

Deploying deep neural networks in real-world settings requires models that are both compact and robust against common corruptions. However, pruning typically degrades corruption robustness, and prior approaches that exploit loss sharpness during training or pruning yield limited gains at deployment-relevant sparsity levels. We propose Magnitude-Weighted Hessian (MWH), an importance score that couples diagonal curvature with parameter magnitude to stabilize Hessian-based pruning at high sparsity. In our experiments, we employ Adaptive Sharpness-Aware Minimization (ASAM) as a common optimizer for all compared methods to control for the effect of loss-surface flattening and to examine its compatibility with pruning. Across CIFAR-10-C and CIFAR-100-C, MWH achieved consistently stronger corruption robustness than existing pruning baselines at 80-90% sparsity, while preserving clean accuracy. We additionally quantified the robustness-throughput trade-off and report practical inference throughput gains at deployment-relevant sparsity levels. Finally, loss landscape diagnostics corroborated that the proposed approach yields flatter minima after pruning.