- Topic: Attacking On Disrupting-Deepfakes: Methodology Transferability And Perturbation Removal
- Designed an auto‑encoder to eliminate disruptions in deepfakes, examining its transferability across various deepfake models
- Validated that the perturbed images had no adverse effects on the performance of other deepfake models, resulting in an MSE score of 0.75
Disrupting Deepfakes: Adversarial Attacks Against Conditional Image Translation Networks and Facial Manipulation Systems
https://arxiv.org/pdf/2003.01279.pdf
Name: Ting-Chih Chen and Xiao Guo
├── Task-1
├── Dataset
├── Readme.md
├── CelebA-100.zip
└── target_images.zip
├── GHOST
├── Readme.md
└── Task1_GHOST.ipynb
├── StyleGAN-NADA
├── stylegan_nada.ipynb
├── Readme.md
├── resultsA.zip
├── resultsK.zip
└── resultsO.zip
├── fewshot-face-translation-GAN
├── Results
├── Results_ariana.zip
├── Results_kobe.zip
└── Results_obama.zip
├── Task1_Online_tool.ipynb
└── Readme.md
├── Evaluation.py
└── Results - Task1.csv
Model-1: fewshot-face-translation-GAN
Model-2: GHOST
Model-3: StyleGAN-NADA
GANs | MSE | SSIM | PSNR |
---|---|---|---|
fewshot-face-translation-GAN | 57.684 | 0.951 | 32.295 |
GHOST | 0.737 | 0.998 | 50.396 |
StyleGAN-NADA | 247.152 | 0.777 | 25.259 |
StyleGAN-Baseline | 226.91 | 0.799 | 26.611 |
├── Task-2
├──Autoencoder.ipynb
├──Restormer.ipynb
└──Task2DataSet.zip
Model-1: Auto-encoder
Model-2: Restormer