Federated Learning for Face Recognition with Gradient Correction

Authors: Yifan Niu, Weihong Deng1999-2007

AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments have been conducted to validate the superiority of Fed GC which can match the performance of conventional centralized methods utilizing full training dataset on several popular benchmark datasets. Extensive experiments and ablation studies have been conducted and demonstrate the superiority of the proposed Fed GC on several popular benchmark datasets.
Researcher Affiliation Academia Beijing University of Posts and Telecommunications nyf@bupt.edu.cn, whdeng@bupt.edu.cn
Pseudocode Yes Algorithm 1: Fed GC. 1: Input. The K clients are indexed by k and hold local data distributes over Sk, η is learning rate. 2: Server initializes model parameters θ0, W 0 3: for each round t = 0, 1, ..., T 1 do 4: Server initializes k-th client model with θt, W t k. 5: for each client k = 1, 2, ..., K do 6: The k-th client computes local Softmax 7: (θt+1 k , W t+1 k ) (θt, W t k) η ℓk xk i , yk i , 8: and sends (θt+1 k , W t+1 k ) to the server. 9: end for 10: Server aggregates the model parameters: 11: θt+1 PK k=1 nk n θt+1 k 12: W t+1 = W t+1 k , . . . , W t+1 K T 13: Server applys gradient correction: 14: W t+1 W t+1 λη W t+1Reg W t+1 15: end for 16: Output. θT , W T
Open Source Code No The paper does not provide an explicit statement about the availability of open-source code or a link to a code repository for the described methodology.
Open Datasets Yes We employ CASIA-Web Face (Yi et al. 2014) as training set.
Dataset Splits No The paper describes dividing the training set into clients for the federated learning setup ('randomly divide training set into 36 clients') and uses various datasets for testing, but it does not specify a distinct validation set split with percentages or sample counts for hyperparameter tuning or early stopping.
Hardware Specification Yes We train the model with 2 synchronized 1080Ti GPUs on Pytorch.
Software Dependencies No The paper mentions 'Pytorch' as a software dependency but does not specify a version number.
Experiment Setup Yes The learning rate is set to a constant of 0.1. The batch size is set as 256. We set momentum as 0.9 and weight decay as 5e-4.