Leveraging Frequency Analysis for Deep Fake Image Recognition
Authors: Joel Frank, Thorsten Eisenhofer, Lea Schönherr, Asja Fischer, Dorothea Kolossa, Thorsten Holz
ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We perform a comprehensive frequency-domain analysis of images generated by various popular GANs, revealing severe artifacts common across different neural network architectures, data sets, and resolutions. We demonstrate the effectiveness of employing frequency representations for detecting GAN-generated deep fake images by an empirical comparison against state-of-the-art approaches. |
| Researcher Affiliation | Academia | 1Ruhr-University Bochum, Horst G ortz Institute for ITSecurity, Bochum, Germany. |
| Pseudocode | No | The paper does not contain any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | The code to reproduce our experiments and plots as well as all pre-trained models are available online at github.com/RUB-Sys Sec/GANDCTAnalysis. |
| Open Datasets | Yes | The images are either from the Flickr-Faces-HQ (FFHQ) data set or from a set generated by Style GAN (Karras et al., 2019). We trained three different versions of Style GAN on the LSUN bedrooms (Yu et al., 2015) dataset. Yu et. al. trained four different GANs (Pro GAN (Karras et al., 2018), SN-DCGAN (Miyato et al., 2018), Cramer GAN (Bellemare et al., 2017), and MMDGAN (Bi nkowski et al., 2018)) on the Celeb A (Liu et al., 2015) and LSUN bedrooms dataset (Yu et al., 2015). |
| Dataset Splits | Yes | We split each set into 10,000 training, 1,000 validation and 5,000 test images, resulting in a training set of 20,000, a validation set of 2,000, and a test set of 10,000 samples. We then partition these samples into 100,000 training, 20,000 validation and 30,000 test images, resulting in a combined set of 500,000 training, 100,000 validation and 150,000 test images. |
| Hardware Specification | Yes | All experiments in this chapter were performed on a server running Ubuntu 18.04, with 192 GB RAM, an Intel Xeon Gold 6230, and four Nvidia Quadro RTX 5000. |
| Software Dependencies | No | The paper mentions the operating system 'Ubuntu 18.04' and the 'Adam optimizer'. However, it does not provide specific version numbers for key software components such as programming languages (e.g., Python version) or libraries (e.g., PyTorch, TensorFlow, CUDA versions). |
| Experiment Setup | Yes | We optimize the linear regression models using the Adam optimizer (Kingma & Ba, 2015) with an initial learning rate of 0.001, minimizing the binary cross-entropy with l2 regularization. We select the regularization factor λ via grid search from λ {10 1, 10 2, 10 3, 10 4} on the validation data, picking the one with the best score. For training our CNN, we use the Adam optimizer with an initial learning rate of 0.001 and a batch size of 1024, minimizing the cross-entropy loss of the model. |