Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
FracFace: Breaking the Visual Clues—Fractal-Based Privacy-Preserving Face Recognition
Authors: Wanying Dai, Beibei Li, Naipeng Dong, Guangdong Bai, Jin Song Dong
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments conducted on multiple public face recognition benchmarks demonstrate that the proposed Frac Face significantly reduces the visual recoverability of facial features, while maintaining high recognition accuracy, as well as the superiorities over state-of-the-art privacy protection approaches. |
| Researcher Affiliation | Academia | Wanying Dai1,4, Beibei Li1 *, Naipeng Dong2 *, Guangdong Bai3, Jin Song Dong4 1Sichuan University 2The University of Queensland 3City University of Hong Kong 4National University of Singapore EMAIL, EMAIL, EMAIL EMAIL, EMAIL |
| Pseudocode | Yes | A.5 The Algorithm of FCR Algorithm 1 Frequency Channel Refining (FCR) A.6 The Algorithm of FFM Algorithm 2 Frequency Fractal Mapping (FFM) |
| Open Source Code | Yes | Code is available at https://anonymous.4open.science/r/Frac Face. |
| Open Datasets | Yes | The model is trained on the MS1Mv2 [9] dataset, which is widespread adoption as a standard benchmark in face recognition ensures fair and consistent comparisons with prior work [27], [29], [28]. To assess privacy robustness, we employ three of the deep learning-based adversaries: a lightweight U-Net [38] autoencoder for reconstruction-based attacks, a Style GAN [18] generator for generative attacks, and a PGDiff[49] generator based on reverse diffusion process. These attackers rigorously test the irreversibility and security of Frac Face transformations. We evaluate on standard benchmarks, including LFW [11], Celeb A [22], Age DB [31], CFP-FP [39], CALFW [55], CPLFW [54], IJB-B [46], and IJB-C [26]. |
| Dataset Splits | Yes | The model is trained on the MS1Mv2 [9] dataset... We evaluate on standard benchmarks, including LFW [11], Celeb A [22], Age DB [31], CFP-FP [39], CALFW [55], CPLFW [54], IJB-B [46], and IJB-C [26]. ... This procedure was uniformly applied across training and evaluation datasets, including MS1M-Arc Face, LFW, and Age DB. |
| Hardware Specification | Yes | We trained the Frac Face model on the MS1Mv2 dataset using Py Torch with two RTX 6000 GPUs (49 GB VRAM each). |
| Software Dependencies | No | We trained the Frac Face model on the MS1Mv2 dataset using Py Torch with two RTX 6000 GPUs (49 GB VRAM each). Training the Frac Face model on MS1Mv2 took about 8 days for a total of 50 epochs. During training, the peak memory usage per GPU was about 45GB. The input comprised 81-channel feature maps produced by the Frac Face pipeline, incorporating DCT transformation, Frequency Channel Refinement (FCR), and Frequency Fractal Mapping (FFM). An IR-50 backbone with Arc Margin loss was employed to enhance identity discrimination. Optimization was performed using Adam W (lr=0.001, weight decay=1e-4) with a cosine annealing scheduler. Training leveraged automatic mixed precision (AMP) and gradient clipping (max norm 5.0) for stability. We monitored performance via Tensor Board and validated on a held-out set after each epoch. Data loading was parallelized with 8 workers and prefetching, and all experiments used torch.backends.cudnn.benchmark=True for optimal GPU performance. |
| Experiment Setup | Yes | Training the Frac Face model on MS1Mv2 took about 8 days for a total of 50 epochs. During training, the peak memory usage per GPU was about 45GB. The input comprised 81-channel feature maps produced by the Frac Face pipeline, incorporating DCT transformation, Frequency Channel Refinement (FCR), and Frequency Fractal Mapping (FFM). An IR-50 backbone with Arc Margin loss was employed to enhance identity discrimination. Optimization was performed using Adam W (lr=0.001, weight decay=1e-4) with a cosine annealing scheduler. Training leveraged automatic mixed precision (AMP) and gradient clipping (max norm 5.0) for stability. |