IEEE Computational Intelligence Magazine - August 2022 - 32

In the following, the average results of 20 run experiments
are collected. For the blockchain setup, the total computation
resource is set to T 200
Sum
=
and tT
for each training node, and the
total number of clients to N = 20. In each communication
round, each client uses tB
time resources to pursue a learning epoch, where tB = 2
for all of the experiments. Letting i = tt /TB, a larger i implies
that the client allots more computing resources to learning in
each communication round.
B. Investigation of Local Differential Privacy
Local differential privacy is applied to each client by adding
random Gaussian noise to the uploaded models in each communication
round. The testing accuracies of the FashionMNIST
and Cifar-10 datasets are plotted in Fig. 2 with respect
to different privacy levels e . In addition, an adaptive noise
decaying method is compared with the constant one, which
will decrease the noise power when the accuracy stops increasing.
The figure further shows that the system achieves higher
performance with a larger value of e, which is under weaker
privacy protection, and the adaptive method can further
improve the learning performance under the same level of privacy
protection.
C. Investigation of Resource Allocation
In this subsection, the resource-allocation results and the training
loss values with different ratios (i) of both datasets are plotted
in Fig. 3. The figure shows the system performance for
different ratios as the number of total communication rounds
time resources to generate a block
increases. Usually, a lower loss function value represents better
training performance. In detail, it can be found that there exists
an optimal total communication round (K) for each computing
ratio i. For example, the smallest training-loss value can be
obtained if clients stop learning in 14 communication rounds
each with 15 learning epochs when x = 1 in the FashionMNIST
dataset. Moreover, for different computing ratios, the
optimal loss value tends to be different. This is due to the fact
that the optimal number of local learning epochs varies according
to different values of i. In addition, similar trends can be
found in the Cifar-10 dataset.
D. Investigation of Lazy Clients
In this subsection, the impact of lazy clients on the proposed
framework is investigated. The signal-to-noise ratio (SNR) is used
to denote the ratio of the power of original model parameters to
2.2
2.4
2.6
1.4
1.6
1.8
2
5101520
Total Communication Rounds
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
= 100
= 10
= 5
= 1
Constant Noise (Fashion-MNIST)
Adaptive Noise (Fashion-MNIST)
Constant Noise (Cifar-10)
Adaptive Noise (Cifar-10)
FIGURE 2 Learning performance with respect to different privacy levels.
(a)
6.5
7
5.5
6
4.5
5
3.5
4
3
Total Communication Rounds
5101520
(b)
θ = 4
θ = 2
θ = 1
θ = 0.5
FIGURE 3 Learning performance of different total communication
rounds under different resource-allocation ratios. (a) Fashion-Mnist
(b) Cifar-10.
32 IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE | AUGUST 2022
Testing Accuracy
Training Loss
Training Loss
IEEE Computational Intelligence Magazine - August 2022

Table of Contents for the Digital Edition of IEEE Computational Intelligence Magazine - August 2022
