Symmetry, Vol. 15, Pages 1074: Enhanced Example Diffusion Model via Style Perturbation

Symmetry, Vol. 15, Pages 1074: Enhanced Example Diffusion Model via Style Perturbation

Symmetry doi: 10.3390/sym15051074

Authors: Haiyan Zhang Guorui Feng

With the extensive applications of neural networks in several fields, research on their security has become a hot topic. The digitization of paintings attracts our interest in the security of artistic style classification tasks. The concept of symmetry is commonly adopted in the construction of deep learning models. However, we find that low-quality artistic examples can fool high-performance deep neural networks. Therefore, we propose the enhanced example diffusion model (EDM) for low-quality paintings to symmetrically generate high-quality enhanced examples with positive style perturbations, which improves the performance of the deep learning-based style classification model. Our proposed framework consists of two parts: a style perturbation network that transforms the inputs into the latent space and extracts style features to form a positive style perturbation, and a conditional latent diffusion model that generates high-quality artistic features. High-quality artistic images are combined with positive style perturbations to generate artistic style-enhanced examples. We conduct extensive experiments on synthetic and real datasets, and find the effectiveness of our approach in improving the performance of deep learning models.