A Closer Look at Fourier Spectrum Discrepancies for CNN-generated Images Detection

CNN-based generative modelling has evolved to produce synthetic images\nindistinguishable from real images in the RGB pixel space. Recent works have\nobserved that CNN-generated images share a systematic shortcoming in\nreplicating high frequency Fourier spectrum decay attributes. Furthermore,\nthese works have successfully exploited this systematic shortcoming to detect\nCNN-generated images reporting up to 99% accuracy across multiple\nstate-of-the-art GAN models.\n In this work, we investigate the validity of assertions claiming that\nCNN-generated images are unable to achieve high frequency spectral decay\nconsistency. We meticulously construct a counterexample space of high frequency\nspectral decay consistent CNN-generated images emerging from our handcrafted\nexperiments using DCGAN, LSGAN, WGAN-GP and StarGAN, where we empirically show\nthat this frequency discrepancy can be avoided by a minor architecture change\nin the last upsampling operation. We subsequently use images from this\ncounterexample space to successfully bypass the recently proposed forensics\ndetector which leverages on high frequency Fourier spectrum decay attributes\nfor CNN-generated image detection.\n Through this study, we show that high frequency Fourier spectrum decay\ndiscrepancies are not inherent characteristics for existing CNN-based\ngenerative models--contrary to the belief of some existing work--, and such\nfeatures are not robust to perform synthetic image detection. Our results\nprompt re-thinking of using high frequency Fourier spectrum decay attributes\nfor CNN-generated image detection. Code and models are available at\nhttps://keshik6.github.io/Fourier-Discrepancies-CNN-Detection/\n