several federated learning algorithms under various
network configurations in order to confirm the
convergence of the model. As shown in Table 1,
taking the MOON algorithm as an example, when
using a CNN network structure, the loss value
gradually decreases from 2.31 to 1.00 with increasing
training rounds. Similar model convergence trends
can be observed in the FedAvg algorithm and DNN
network structure. It should be noted that the model
can converge with only about 10 epochs, indicating
that the distributed model training concept of
federated learning can improve the learning
efficiency of the model.
Table 1: Comparison of training loss for various models.
Epoch
MOON
_cnn
FedAvg
_cnn
MOON
_dnn
FedAvg
_dnn
1 2.3128 2.3128 6.2129 6.2129
2 2.1526 2.1518 2.4193 2.4300
3 1.8138 1.8079 1.8428 1.8334
4 1.5184 1.5178 1.5930 1.5909
5 1.3397 1.3419 1.4467 1.4476
6 1.2336 1.2395 1.3318 1.3350
7 1.1611 1.1661 1.2595 1.2571
8 1.1114 1.1121 1.2034 1.2004
9 1.0700 1.0634 1.1543 1.1564
10 1.0402 1.0410 1.1211 1.1213
11 1.0053 1.0044 1.0954 1.0969
In addition, an additional set of experiments was
conducted to compare the recognition accuracy of
different algorithms, and Figure 3 shows the results.
As the training process increases, the average testing
accuracy of MOON_cnn on the Fashion MINIST
dataset can be improved from 0.21 to 0.63. In the
early stages of training (0-3 epochs), there are some
differences in the testing accuracy of different
network structures and algorithms, and MOON_cnn
shows the best results. As the model is further fully
trained, the FedAvg algorithm and MOON algorithm
will converge to approximately the same accuracy.
Figure 3: Average test accuracy of different models.
5 CONCLUSIONS
To explore the improvement effect of federated
learning on image recognition tasks, this paper
designs a unified client and server based on the
PFLlib platform, integrating representative federated
learning algorithms with different network structures.
Numerous trials have verified the effectiveness of the
work in this paper. The module partitioning and
overall structural design of the PFLlib platform
showed significant clarity in the experiment. By
storing the client and server implementations of
different algorithms in separate folders and adopting
the inheritance mechanism of base classes, the
platform maintains consistency and maintainability.
REFERENCES
Baldini, I., Castro, P., Chang, K., Cheng, P., Fink, S., 2017.
Serverless computing: Current trends and open
problems. In Research advances in cloud computing, 1-
20.
Cheng, K., Fan, T., Jin, Y., Liu, Y., Chen, T., Papadopoulos,
D., Yang, Q., 2021. Secureboost: A lossless federated
learning framework. IEEE Intelligent Systems, 2021,
36(6):87-98.
Durmus, A., E., A., Zhao, Y., Ramon, M., N., Matthew, M.,
Paul, N., W., Venkatesh, S., 2021. Federated learning
based on dynamic regularization. In arxiv preprint
arXiv:2111.04263v2.
Li, Q., He, B., Dawn, S., 2021. Model-Contrastive
Federated Learning. In Proceeding of IEEE/CVF
Conference on Computer Vision and Pattern
Recognition (CVPR).
Liu, R., 2020. Fedsel: Federated sgd under local differential
privacy with top-k dimension selection. Database
Systems for Advanced Applications: 25th International
Conference, DASFAA 2020, Jeju, South Korea,
September 24-27, 2020, Proceedings, Part I 25.
Springer International Publishing, 2020.
McMahan, H., B., Eider, M., Daniel, R., Seth, H., Blaise,
A., y., A., 2017. Communication-Efficient Learning of
Deep Networks from Decentralized Data. In
Proceedings of the 20th International Conference on
Artificial Intelligence and Statistics 2017.
Sun, H., Lang, W., Xu, C., Liu, N., Zhou, H., 2023. Graph-
based discriminative features learning for fine-grained
image retrieval. Signal Processing: Image
Communication, 110:116885.
Wang, J., Wang, Y., Liu, L., Yin, H., Ye, N., Xu, C., 2023.
Weakly Supervised Forest Fire Segmentation in UAV
Imagery Based on Foreground-Aware Pooling and
Context-Aware Loss. Remote Sensing. 15, no.14: 3606.
Xu, C., Jiang, H., Peter, Y., Khan, Z., A., Chen, Y., 2020.
MHW-PD: A robust rice panicles counting algorithm