The Role of Noisy Data in Improving CNN Robustness for Image Classification

TL;DR

The paper investigates whether deliberately injecting controlled noise into CNN training data improves image-classification robustness to real-world degradations. Using CIFAR-10 and a ResNet-18 baseline, it shows that exposing models to modest levels of noise (as low as around 5–10% of training samples) can significantly reduce test loss and improve accuracy on corrupted test sets with only minor or manageable impacts on clean-set performance. The approach acts as a simple, effective regularizer, offering a practical trade-off between data cleanliness and real-world resilience. These findings advocate reevaluating traditional data-quality assumptions and suggest noise-augmented training as a viable strategy for deploying robust vision systems in imperfect environments.

Abstract

Data quality plays a central role in the performance and robustness of convolutional neural networks (CNNs) for image classification. While high-quality data is often preferred for training, real-world inputs are frequently affected by noise and other distortions. This paper investigates the effect of deliberately introducing controlled noise into the training data to improve model robustness. Using the CIFAR-10 dataset, we evaluate the impact of three common corruptions, namely Gaussian noise, Salt-and-Pepper noise, and Gaussian blur at varying intensities and training set pollution levels. Experiments using a Resnet-18 model reveal that incorporating just 10\% noisy data during training is sufficient to significantly reduce test loss and enhance accuracy under fully corrupted test conditions, with minimal impact on clean-data performance. These findings suggest that strategic exposure to noise can act as a simple yet effective regularizer, offering a practical trade-off between traditional data cleanliness and real-world resilience.

Figures (11)

• Figure 1: Impact of different types and intensities of image degradation on CIFAR-10.
• Figure 2: Average test loss across different Gaussian noise intensities and varying training noise percentages. Shaded areas indicate $\pm 1$ standard deviation over 10 runs.
• Figure 3: Average test loss across different Salt-and-Pepper noise intensities and varying training noise percentages. Shaded areas indicate $\pm 1$ standard deviation over 10 runs.
• Figure 4: Average test loss across different Gaussian blur intensities and varying training noise percentages. Shaded areas indicate $\pm 1$ standard deviation over 10 runs.
• Figure 5: Average test accuracy across different Gaussian noise intensities and varying training noise percentages. Shaded areas indicate $\pm 1$ standard deviation over 10 runs.