Classificação de imagens reais e sintéticas com transformada de Fourier e modelos de deep learning

Abstract

Resumo: Com os avanços recentes dos modelos de difusão (Diffusion Models – DM) aplicados à geração automática de imagens, tornou-se cada vez mais desafiador diferenciar imagens reais daquelas criadas artificialmente. Este estudo propõe a construção de classificador capaz de realizar essa distinção, com base na análise espectral obtida por meio da Transformada de Fourier (FFT) e de modelos de Deep Learning. A proposta se apoia no trabalho de Bammey [1], que demonstrou como os picos de magnitude da FFT podem funcionar como indicadores eficazes para diferenciar imagens. Foram desenvolvidas abordagens alternativas que incluem o uso de filtros passa-alta, compressão espectral com escalas logarítmica e fracionária, aumento do número de picos de magnitude e de fase espectral, uso de redes neurais convolucionais e de arquitetura transformer, Vision Transformer (ViT). Os modelos HistGradientBoostingClassifier, MLP (Multilayer Perceptron) de até 2 camadas e rede neural convolucional simples baseada na LeNet-5 nomeada de SysLeNet2025 foram treinados e testados. Os resultados apontaram que as variações introduzidas melhoraram o desempenho da classificação, evidenciando o potencial da análise espectral como recurso relevante na identificação de imagens sintéticas

Abstract: With the recent advances of diffusion models (Diffusion Models – DM) applied to automatic image generation, it has become increasingly challenging to distinguish real images from those created artificially. This study proposes the development of a classifier capable of making this distinction, based on spectral analysis obtained through the Fourier Transform (FFT) and Deep Learning models The proposal is supported by the work of Bammey [1], who demonstrated how FFT magnitude peaks can serve as effective indicators for differentiating images. Alternative approaches were developed, including the use of high-pass filters, spectral compression with logarithmic and fractional scales, an increased number of magnitude and phase peaks, and the use of convolutional neural networks and transformer architectures, such as the Vision Transformer (ViT). The models HistGradientBoostingClassifier, MLP (Multilayer Perceptron) with up to 2 layers, and a simple convolutional neural network based on LeNet-5, named SysLeNet2025, were trained and evaluated. The results showed that the introduced variations improved classification performance, highlighting the potential of spectral analysis as a valuable resource for identifying synthetic images. The results showed that the introduced variations improved classification performance, highlighting the potential of spectral analysis as a valuable resource for identifying synthetic images