Generating Synthetic Data Using 2D Multifractal Detrended Fluctuation Analysis Generative Adversarial Networks

Abstract

A novel 2-dimensional Multifractal Detrended Fluctuation Analysis (2D-MFDFA) as a discriminator for Generative Adversarial Networks (GAN) is presented in this work. The proposed approach demonstrates improved generated data quality at early iterations on the MNIST dataset. This algorithm produces synthetic images rapidly while maintaining lower computational complexity compared to existing GAN-based methods. The effectiveness and performance of this methodology are evaluated using the Fréchet Inception Distance (FID) score. The proposed 2D-MFDFA GAN is deployed on the AMD Xilinx Kintex7 FPGA KC705 using the NNGen framework, with validation conducted via a hardware experimental setup for power efficiency and performance. The experimental results reveal that the proposed method achieves an 85% improvement in FID score, reduces power consumption by 81%, and operates 94% faster, significantly outperforming comparable GAN models reported in the literature.