Novel Image Data Augmentation Technique for Deep Learning Using Least Significant Bit Encryption
Rawan Ghnemat, Sami Almashaqbeh · 2024
Deep convolutional neural networks rely heavily on large datasets for training and testing in order to avoid overfitting. However, in fields like medical imaging analysis, big data may not be available. To address this issue, various techniques can be employed to prevent overfitting, such as data augmentation. Data augmentation involves creating new versions of existing photos using different tools like cropping, and flipping. By applying these methods, new images with variations can be generated, which can help machine learning models perform better while retaining the same underlying information as the original image. The focus of our study is on creating a novel approach to enhance image data augmentation by producing new versions of images that are similar enough to the original to be useful for developing machine learning models while preserving the labels. Specifically, our method involves modifying the least significant bit (LSB) of RGB images. Results show that models trained using StegAug significantly outperform models trained using other CNN architectures, proving the applicability of our technique. Using ResNet-20, ResNet-110, and ResNet-44, our strategy increases the accuracy for CIFAR-10 by 0.87%, 0.45%, and 0.60%, respectively. Applying our method using 2% of the CIFAR-10 subset increases the base accuracy by 0.488 using ResNet-20, 21.22 using ResNet-110, and 12.73 using ResNet-44. As the depth increases and we delve into very deep learning models, the need for a larger training dataset also increases. This proves that our proposed technique is very effective for augmenting image data in the context of very deep learning networks.