CAMOUFLaGE: Controllable AnoniMizatiOn throUgh diFfusion-based image coLlection GEneration

Code Here

Official implementations of "Latent Diffusion Models for Attribute-Preserving Image Anonymization" and "Harnessing Foundation Models for Image Anonymization".

Latent Diffusion Models for Attribute-Preserving Image Anonymization

[Paper]

This paper presents, to the best of our knowledge, the first approach to image anonymization based on Latent Diffusion Models (LDMs). Every element of a scene is maintained to convey the same meaning, yet manipulated in a way that makes re-identification difficult. We propose two LDMs for this purpose:

CAMOUFLaGE-Base
CAMOFULaGE-Light

The former solution achieves superior performance on most metrics and benchmarks, while the latter cuts the inference time in half at the cost of fine-tuning a lightweight module.

Compared to state-of-the-art, we anonymize complex scenes by introducing variations in the faces, bodies, and background elements.

CAMOUFLaGE-Base

CAMOUFLaGE-Base exploits a combination of pre-trained ControlNets and introduces an anonymizazion guidance based on the original image.

More details on its usage can be found here.

CAMOUFLaGE-Light

CAMOUFLaGE-Light trains a lightweight IP-Adapter to encode key elements of the scene and facial attributes of each person.

More details on its usage can be found here.

Harnessing Foundation Models for Image Anonymization

[Paper]

We explore how foundation models can be leveraged to solve tasks, specifically focusing on anonymization, without the requirement for training or fine-tuning. By bypassing traditional pipelines, we demonstrate the efficiency and effectiveness of this approach in achieving anonymization objectives directly from the foundation model’s inherent knowledge.

CAMOUFLaGE-Face

We examine how foundation models can generate anonymized images directly from textual descriptions. Two models were employed for information extraction: FACER, used to identify the 40 CelebA-HQ attributes, and DeepFace, used to determine ethnicity and age. Using this rich information, we craft captions to guide the generation process. Classifier-free guidance was employed to push the image content in the direction of the positive prompt P and far from the negative prompt ¬P.

More details on its usage can be found here.

Citation

If you find CAMOUFLaGE-Base and/or CAMOUFLaGE-Light useful, please cite:

@misc{camouflage,
title={Latent Diffusion Models for Attribute-Preserving Image Anonymization}, 
author={Luca Piano and Pietro Basci and Fabrizio Lamberti and Lia Morra},
year={2024},
eprint={2403.14790},
archivePrefix={arXiv},
primaryClass={cs.CV}
}

If you find CAMOUFLaGE-Face useful, please cite:

@inproceedings{pianoGEM24,
title={Harnessing Foundation Models for Image Anonymization},
author={Piano, Luca and Basci, Pietro and Lamberti, Fabrizio and Morra, Lia},
booktitle={2024 IEEE CTSoc Gaming, Entertainment and Media},
year={2024},
organization={IEEE}
}