Adversarial Robust Safeguard for Evading Deep Facial Manipulation

Authors: Jiazhi Guan, Yi Zhao, Zhuoer Xu, Changhua Meng, Ke Xu, Youjian Zhao

AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental To evaluate it, we conduct experiments on four manipulation methods and compare recent works comprehensively. The results of our method exhibit good visual effects with pronounced robustness against varied perturbations at different levels.
Researcher Affiliation Collaboration Jiazhi Guan1, Yi Zhao2*, Zhuoer Xu3, Changhua Meng3, Ke Xu1,4, Youjian Zhao1,4* 1DCST, BNRist, Tsinghua University 2Beijing Institute of Technology 3Ant Group 4Zhongguancun Laboratory
Pseudocode No The paper provides a diagram of the proposed model (Figure 2) and describes its components, but it does not include any formal pseudocode or algorithm blocks.
Open Source Code No The paper states, 'For all the manipulation methods, we collect the open-source codes and model weights from their official implementations,' referring to baselines, but it does not provide any statement or link for the open-sourcing of its own proposed method's code.
Open Datasets Yes For p Sp-mix, p Sp-recon, and Style Clip, we randomly select 10000 images from FFHQ (Karras, Laine, and Aila 2019) in training, 2000 images for validation, and 2000 images for testing. For Sim Swap, we also introduce a custom subset of VGGFace2 (Cao et al. 2018) including 10000, 2000, and 2000 images for train, validation, and test, respectively.
Dataset Splits Yes For p Sp-mix, p Sp-recon, and Style Clip, we randomly select 10000 images from FFHQ (Karras, Laine, and Aila 2019) in training, 2000 images for validation, and 2000 images for testing. For Sim Swap, we also introduce a custom subset of VGGFace2 (Cao et al. 2018) including 10000, 2000, and 2000 images for train, validation, and test, respectively.
Hardware Specification No The paper mentions running methods 'with the same experimental environment' but does not specify any hardware details such as GPU models, CPU types, or memory.
Software Dependencies No The paper specifies that functions are implemented using U-net architecture and Conv-BN-ReLU layers, but it does not provide specific version numbers for any software dependencies like Python, PyTorch/TensorFlow, or other libraries.
Experiment Setup Yes ϵ is set to 0.1 by default. The channel C for noise encoding is 128. We use Adam optimizer with a learning rate of 10 4 in training. And our model converges at around 30 epochs.