Yuan Gan · Jiaxu Miao · Yunze Wang . Yi Yang

Abstract: Advances in talking-head animation based on Latent Diffusion Models (LDM) enable the creation of highly realistic, synchronized videos. These fabricated videos are indistinguishable from real ones, increasing the risk of potential misuse for scams, political manipulation, and misinformation. Hence, addressing these ethical concerns has become a pressing issue in AI security. Recent proactive defense studies focused on countering LDM-based models by adding perturbations to portraits. However, these methods are ineffective at protecting reference portraits from advanced image-to-video animation. The limitations are twofold: 1) they fail to prevent images from being manipulated by audio signals, and 2) diffusion-based purification techniques can effectively eliminate protective perturbations. To address these challenges, we propose Silencer, a two-stage method designed to proactively protect the privacy of portraits. First, a nullifying loss is proposed to ignore audio control in talking-head generation. Second, we apply anti-purification loss in LDM to optimize the inverted latent feature to generate robust perturbations. Extensive experiments demonstrate the effectiveness of Silencer in proactively protecting portrait privacy. We hope this work will raise awareness among the AI security community regarding critical ethical issues related to talking-head generation techniques.

cd Silencer-I
conda create -n silencer python=3.10
conda activate silencer
pip install -r requirements.txt

Download the pretrained models from hallo.

Download our preprocessed CelebA-HQ and TH1KH datasets from Yandex.

These files should be organized as follows:

./Silencer-I/
|-- ...
|-- pretrained_models/
|   |-- ...
|-- th1kh/
|   |-- ...
|-- celebahq_512_dataset/
|   |-- ...

Change the path of pretrained_models into your own absolute path in ./Silencer-I/build/lib/hallo/utils/util.py and ./Silencer-I/configs/train/stage2.yaml. Then install Hallo by executing the following command in Silencer-I.

pip install .

PS: It may change the version of protobuf. You need to keep the protobuf==3.20.3 with pip install protobuf==3.20.3.

In Silencer-I, you can run the following command to proactively protect images in CelebA-HQ and TH1KH datasets.

CUDA_VISIBLE_DEVICES=0 python protect/protect_hallo.py attack.img_path='celebahq_512_dataset/celebahq_512' \
    attack.output_path='protect/out_celebahq_512/' attack.mode='hallo' attack.g_mode='-' \
    attack.min_timesteps=200
CUDA_VISIBLE_DEVICES=0 python protect/protect_hallo.py attack.img_path='th1kh/th1kh_imgs_100' \
    attack.output_path='protect/out_th1kh_512/' attack.mode='hallo' attack.g_mode='-' \
    attack.min_timesteps=200

Refer to test_hallo_celebahq_hallo-.py and test_hallo_th1kh_hallo-.py.

Useage examples are in test_hallo_celebahq.sh and test_hallo_th1kh.sh.

For example:

python test_hallo_th1kh_hallo-.py 0 0 &
python test_hallo_th1kh_hallo-.py 1 0 &
python test_hallo_th1kh_hallo-.py 2 0 &
python test_hallo_th1kh_hallo-.py 3 0 &

Please refer to the LRW evaluation process of EAT here.

1. Generate ground truth videos with original portraits and the hallo model.
2. Generate ground truth for evaluation with GT videos.
3. Evaluate the generated videos with protected portraits.

Based on the silencer, you need to install additional packages from DiffAttack.

You can refer to my environment here. Untar the file into the envs folder of conda/mamba, then install hallo and protobuf.

Download the attack code from here and untar it into Silencer-II.

In Silencer-II/DiffAttack, change the root_hallo and the ouput path of Silencer-I into your own path in main_hallo_attnloss.py.

Then you can run the following command for protection:

CUDA_VISIBLE_DEVICES=0 FORCE_MEM_EFFICIENT_ATTN=1 python  main_hallo_attnloss.py --model_name resnet \
    --save_dir ./adv_out/celebahq_adamw_iter200_1e-2_hallo1_512_19_+mse_t200_s100_fmask_10_100 \
    --images_root ../../Silencer-I/celebahq_512_dataset/celebahq_512/ --attack_mode 'hallo' --g_mode '-'\
    --res 512 --iterations 200 --attack_loss_weight 1 --attack_mse_loss_weight 10 \
    --cross_attn_loss_weight 0 --self_attn_loss_weight 0 --start_step 19 \
    --dataset_name celebahq --use_face_mask 1 \
    --index_use_face_mask 100 --part [0-3]
CUDA_VISIBLE_DEVICES=0 FORCE_MEM_EFFICIENT_ATTN=1 python  main_hallo_attnloss.py --model_name resnet \
    --save_dir ./adv_out/th1kh_adamw_iter200_1e-2_hallo1_512_19_+mse_t200_s100_fmask_10_100_20251217   \
    --images_root ../../Silencer-I/th1kh/th1kh_imgs_100/ --attack_mode 'hallo' --g_mode '-'\
    --res 512 --iterations 200 --attack_loss_weight 1 --attack_mse_loss_weight 10  \
    --cross_attn_loss_weight 0 --self_attn_loss_weight 0 --start_step 19 \
    --dataset_name th1kh --use_face_mask 1 \
    --index_use_face_mask 100 --part [0-3]

Refer to test_hallo_celebahq_diff_hallo-.py and test_hallo_th1kh_diff_hallo-.py.

Usage examples are the same as Silencer-I:

python test_hallo_th1kh_diff_hallo-.py 0 0 &
python test_hallo_th1kh_diff_hallo-.py 1 0 &
python test_hallo_th1kh_diff_hallo-.py 2 0 &
python test_hallo_th1kh_diff_hallo-.py 3 0 &

If you find our work useful, please cite it as follows:

@inproceedings{gan2025silence,
  title={Silence is Golden: Leveraging Adversarial Examples to Nullify Audio Control in LDM-based Talking-Head Generation},
  author={Gan, Yuan and Miao, Jiaxu and Wang, Yunze and Yang, Yi},
  booktitle={Proceedings of the Computer Vision and Pattern Recognition Conference},
  pages={13434--13444},
  year={2025}
}

We acknowledge these works for their public code: Diff-Protect, DiffAttack, Hallo, and so on.