Balancing Privacy and Performance: A Many-in-One Approach for Image Anonymization

Authors: Xuemei Jia, Jiawei Du, Hui Wei, Ruinian Xue, Zheng Wang, Hongyuan Zhu, Jun Chen

AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on identity identification tasks demonstrate that FRO outperforms previous state-of-the-art methods, not only in utility performance but also in visual anonymization.
Researcher Affiliation Academia 1 National Engineering Research Center for Multimedia Software, School of Computer Science, Wuhan University, China 2 Hubei Key Laboratory of Multimedia and Network Communication Engineering, Wuhan University, China 3 Centre for Frontier AI Research (CFAR) & Institute of High Performance Computing (IHPC), A*STAR, Singapore 4 Centre for Frontier AI Research (CFAR) & Institute for Infocomm Research (I2R), A*STAR, Singapore
Pseudocode Yes Algorithm 1: Feature Resembling Algorithm
Open Source Code No The paper does not provide an explicit statement about releasing source code for the described methodology, nor does it include a link to a code repository.
Open Datasets Yes Datasets. MARKET1501 (Zheng et al. 2015) is a largescale re-id benchmark comprising 32,668 images of 1,501 pedestrians, 751 for training and 750 for testing, captured by six cameras. DUKEMTMC-REID (Zheng, Zheng, and Yang 2017) dataset contains 36,441 images of 1,812 persons captured by eight cameras, 702 identities are used as the training set, and 702 persons are used as the query and gallery, respectively. A subset of CASIA (Yi et al. 2014) containing 30,726 face images of 952 identities are selected in our experiments. And LFW (Huang et al. 2008), CFP-FF and CFP-FP (Sengupta et al. 2016), and AGEDB (Moschoglou et al. 2017) are used for downstream face verification.
Dataset Splits Yes MARKET1501 (Zheng et al. 2015) is a largescale re-id benchmark comprising 32,668 images of 1,501 pedestrians, 751 for training and 750 for testing, captured by six cameras. DUKEMTMC-REID (Zheng, Zheng, and Yang 2017) dataset contains 36,441 images of 1,812 persons captured by eight cameras, 702 identities are used as the training set, and 702 persons are used as the query and gallery, respectively.
Hardware Specification Yes The models are trained with one NVIDIA GeForce RTX 4090 GPU using Pytorch.
Software Dependencies No The paper mentions 'Pytorch' but does not specify its version number or any other software dependencies with version numbers.
Experiment Setup Yes Training Details. The models are trained with one NVIDIA GeForce RTX 4090 GPU using Pytorch. We use Res Net-50 (He et al. 2016) as the backbone with Adam Optimizer. The input images are resized to 256 128. The mini-batch size is set to 64, containing 32 persons with 4 images each. The initial learning rate is 3e 4 and is reduced by following an exponentially decaying training schedule. α is set as 0.05.