ARE GAN GENERATED IMAGES EASY TO DETECT? A CRITICAL ANALYSIS OF THE STATE-OF-THE-ART 论文阅读

date

Oct 4, 2023

Last edited time

Oct 3, 2023 06:31 AM

status

Published

slug

ARE GAN GENERATED IMAGES EASY TO DETECT? A CRITICAL ANALYSIS OF THE STATE-OF-THE-ART 论文阅读

tags

summary

type

Post

origin

https://www.notion.so/lazurite/ARE-GAN-GENERATED-IMAGES-EASY-TO-DETECT-A-CRITICAL-ANALYSIS-OF-THE-STATE-OF-THE-ART-e5a80ea2d98346418ae9d639ce1418cb?pvs=4

Field

Plat

Are GAN generated images easy to detect? A critical analysis of...

The advent of deep learning has brought a significant improvement in the quality of generated media. However, with the increased level of photorealism, synthetic media are becoming hardly...

https://arxiv.org/abs/2104.02617

Gragnaniello 等 - 2021 - Are GAN generated images easy to detect A critica.pdf

3386.8KB

GitHub - grip-unina/GANimageDetection

Contribute to grip-unina/GANimageDetection development by creating an account on GitHub.

https://github.com/grip-unina/GANimageDetection

在 Wang2020 框架上，我们引入了以下变化：删除 Imagenet 预训练（无预训练），包括用于残差提取的初始层（残差），按照[28]的建议不要在第一层执行下采样（无-down），通过添加高斯噪声、几何变换、剪切以及亮度和对比度变化来执行更强的增强（strong-aug）。此外，对于no-down变体，我们还更改了主干网络，将ResNet50替换为XceptionNet（Xception no-down）和Efficient-B4（Efficient no-down）。

我们发现避免在架构的第一个块中进行下采样可以使精度平均提高约 15%，就 Pd@5% 而言平均提高 14%。