The official pytorch implementation of "SecureReID: Privacy-Preserving Anonymization for Person Re-Identification" (TIFS 2024).
SecureReID: Privacy-Preserving Anonymization for Person Re-Identification
Mang Ye, Wei Shen, Junwu Zhang, Yao Yang, Bo Du
Wuhan University
Abstract Anonymization methods have gained widespread use in safeguarding privacy. However, conventional anonymization solutions inevitably lead to the loss of semantic information, resulting in limited data utility. Besides, existing deep learning-based anonymization strategies inadvertently alter the identities of pedestrians, rendering them unsuitable for re-identification (Re-ID) tasks. Beyond these limitations, we propose a joint learning reversible anonymization framework that can reversibly generate full-body anonymized images with little performance drop on Re-ID tasks. Despite these advancements, we reveal that the anonymization methods are vulnerable to model attacks, where attackers can utilize the anonymization model and public data to perform recovery and Re-ID tasks on anonymized images. To defend against the potential attack, we introduce the identity-specific encrypt-decrypt (ISED) architecture for enhanced security, where the anonymized images are encrypted using the specific key for each identity. It renders the images computationally inaccessible to attackers while allowing for seamless reversal without loss using the corresponding keys. Extensive experiments demonstrate that the anonymization framework can guarantee Re-ID performance while protecting pedestrian privacy. In addition, we provide both empirical and theoretical evidence to demonstrate the feasibility of model attacks and the effectiveness of our ISED strategy.
2024/01/30 Accepted by IEEE Transactions on Information Forensics and Security in 2024 (TIFS 2024).
2022/11/21 We have released the official codes.
Install
- Clone repo and install requirements.txt in a Python>=3.7.10 environment, including PyTorch>=1.8.
git clone https://github.com/shentt67/SecureReID.git
cd SecureReID
pip install -r requirements.txtPrepare Datasets
- Download Market1501, MSMT17 and CUHK03 datasets from http://www.liangzheng.org/Project/project_reid.html to <path_to_root_of_datasets>.
- Extract dataset and rename to
market1501. The data structure would like:
<path_to_root_of_datasets/market1501>
market1501
bounding_box_test/
bounding_box_train/
query/
- Split Market1501 training set into a smaller training set (80%) and a validation set (20%). Furthermore, the validation set is split into a validation gallery set (80%) and a validation query set (20%).
python3 data/train_split.py --dataset_root_dir '<path_to_root_of_datasets>' --dataset_name 'market1501'- The data structure would like:
<path_to_root_of_datasets/market1501_val>
market1501
bounding_box_test/
bounding_box_train/
query/
bounding_box_val_gallery/
bounding_box_val_query/
- MSMT17 can be split by changing 'market1501' to 'msmt17', CUHK03 can be split using Matlab.
Notes
- Replace <path_to_root_of_datasets> by your own path to the root of datasets, e.g., /data/Dataset.
- For other datasets, replace 'market1501' by 'msmt17' or 'cuhk03'.
- For other desensitization methods, INPUT.TYPE and INPUT.RADIUS can be changed from mosaic 24.0 to blur 12.0 or noise 0.5.
- Jointly training with upgradation, replace <R1_update> by the rank-1 accuracy constraint. When the rank-1 accuracy is higher than the constraint plus a small negative value (set to -0.05) and the privacy constraint is satisfied simultaneously, the supervision will be substituted by the current learned anonymized images.
Baseline
- Train our re-id baseline on hybrid images (raw and desensitized):
python3 tools/main.py --config_file='configs/AGW_baseline.yml' DATASETS.ROOT_DIR "<path_to_root_of_datasets>" DATASETS.NAMES "('market1501_val')" INPUT.TYPE "mosaic" INPUT.RADIUS "24.0" OUTPUT_DIR "('./log/market1501/market1501-base-mosaic')" MODEL.MODE "C"Our Model w/o Supervision Upgradation
- Jointly train our models (Anonymization, Re-ID and Recovery) without upgradation:
python3 tools/main.py --config_file='configs/AGW_baseline.yml' DATASETS.ROOT_DIR "<path_to_root_of_datasets>" DATASETS.NAMES "('market1501_val')" INPUT.TYPE "mosaic" INPUT.RADIUS "24.0" OUTPUT_DIR "('./log/market1501/market1501-wosu-mosaic')"Our Model w/ Supervision Upgradation
- Jointly train our models (Anonymization, Re-ID and Recovery) with upgradation:
python3 tools/main.py --config_file='configs/AGW_baseline.yml' DATASETS.ROOT_DIR "<path_to_root_of_datasets>" DATASETS.NAMES "('market1501_val')" INPUT.TYPE "mosaic" INPUT.RADIUS "24.0" OUTPUT_DIR "('./log/market1501/market1501-full-mosaic')" MODEL.VAL_R1 "<R1_update>"The testing command is just the incremental version of the training command, which is prefixed with the training command and suffixed with:
- TEST.EVALUATE_ONLY is set as 'on' to indicate performing test.
- MODEL.PRETRAIN_CHOICE is set as 'imagenet' or 'self', refer using pretrained resnet50 backbone or pretrained re-id baseline respectively. Set as 'self' in testing.
- MODEL.PRETRAIN_DIR is set as the directory of model to be evaluated.
Baseline
python3 tools/main.py --config_file='configs/AGW_baseline.yml' DATASETS.ROOT_DIR "<path_to_root_of_datasets>" DATASETS.NAMES "('market1501')" INPUT.TYPE "mosaic" INPUT.RADIUS "24.0" OUTPUT_DIR "('./log/market1501/test/market1501-base-mosaic')" MODEL.MODE "C" TEST.EVALUATE_ONLY "('on')" MODEL.PRETRAIN_CHOICE "('self')" MODEL.PRETRAIN_DIR "./log/market1501/market1501-base-mosaic"Our Model w/o Supervision Upgradation
python3 tools/main.py --config_file='configs/AGW_baseline.yml' DATASETS.ROOT_DIR "<path_to_root_of_datasets>" DATASETS.NAMES "('market1501')" INPUT.TYPE "mosaic" INPUT.RADIUS "24.0" OUTPUT_DIR "('./log/market1501/test/market1501-wosu-mosaic')" TEST.EVALUATE_ONLY "('on')" MODEL.PRETRAIN_CHOICE "('self')" MODEL.PRETRAIN_DIR "./log/market1501/market1501-wosu-mosaic"Our Model w/ Supervision Upgradation
python3 tools/main.py --config_file='configs/AGW_baseline.yml' DATASETS.ROOT_DIR "<path_to_root_of_datasets>" DATASETS.NAMES "('market1501')" INPUT.TYPE "mosaic" INPUT.RADIUS "24.0" OUTPUT_DIR "('./log/market1501/test/market1501-full-mosaic')" TEST.EVALUATE_ONLY "('on')" MODEL.PRETRAIN_CHOICE "('self')" MODEL.PRETRAIN_DIR "./log/market1501/market1501-full-mosaic"Pretrained models will be published later.
- Qualitative Results
Qualitative results. (a)/(b)/(c) are desensitized/anonymized/encrypted images. Encrypted images achieve a strong visual confusion effect.
-
Quantitative Results of Re-ID
-
The four values separated by '/' are corresponding to the rank-1 accuracy under raw/crossed/protected settings in the main paper.
-
Market-1501 Dataset
Blur Mosaic Noise Base 84.8/40.1/18.3/67.3 87.4/75.3/70.9/64.3 75.9/50.8/41.7/68.7 Ours w/o U 83.1/68.3/46.8/75.8 86.3/80.1/75.1/73.2 90.4/77.5/67.5/84.4 Ours 91.6/88.2/82.5/89.2 89.4/88.5/86.8/87.0 91.7/83.5/83.8/91.2 -
MSMT17 Dataset
Blur Mosaic Noise Base 30.5/21.3/16.2/15.2 25.0/16.3/14.6/10.6 24.1/20.1/18.5/18.2 Ours w/o U 43.6/28.4/10.0/14.7 34.3/26.2/24.6/20.5 41.9/36.0/29.4/31.3 Ours 51.5/51.1/50.5/48.7 48.6/49.1/48.5/48.1 46.9/48.1/46.2/46.0 -
CUHK03 Dataset
Blur Mosaic Noise Base 30.4/14.6/10.4/8.2 28.5/17.7/15.1/8.8 14.0/9.1/8.6/8.1 Ours w/o U 31.6/13.8/8.9/14.3 24.9/24.2/19.6/12.1 28.1/30.4/30.4/31.9 Ours 41.9/39.2/35.3/33.2 38.8/37.8/30.4/27.6 35.8/36.4/37.9/41.9
-
-
Please cite our work in your publications if it helps your research:
@inproceedings{zhang2022learnable,
title={Learnable privacy-preserving anonymization for pedestrian images},
author={Zhang, Junwu and Ye, Mang and Yang, Yao},
booktitle={Proceedings of the 30th ACM International Conference on Multimedia},
pages={7300--7308},
year={2022}
}
@article{ye2024securereid,
title={SecureReID: Privacy-Preserving Anonymization for Person Re-Identification},
author={Ye, Mang and Shen, Wei and Zhang, Junwu and Yang, Yao and Du, Bo},
journal={IEEE Transactions on Information Forensics and Security},
year={2024}
}
Contact: yemang@whu.edu.cn, weishen@whu.edu.cn