Frequency-Robust Dataset Pruning for Edge and IoT Applications Under Noisy Conditions

In real-world deployment scenarios such as IoT and edge devices, neural networks frequently encounter noisy data that degrades model performance. While existing dataset pruning methods aim to reduce data volume by selecting representative samples, they often overlook the vulnerability of selected samples to noise perturbations. In this paper, we propose a novel noiserobust dataset pruning framework that explicitly considers the frequency characteristics of samples to enhance robustness under noisy conditions. Our method first introduces a multi-step frequency robustness metric based on SmoothGrad and 2D Fourier analysis to quantify the model's attention bias toward high- or low-frequency components throughout training. Samples that rely heavily on high-frequency details-often susceptible to Gaussian noise-are excluded. To preserve diversity and generalizability, we extract semantic embeddings using a pretrained network and perform diversity-aware clustering to select the most informative samples from each cluster. Experiments on CIFAR-10, CIFAR-100, and ImageNet-1K demonstrate that our method consistently outperforms state-of-the-art pruning algorithms across a range of pruning ratios and noise levels, achieving better performance under realistic deployment constraints. Our approach offers a principled and scalable solution for constructing compact yet noise-resilient training subsets.