IEEE Computational Intelligence Magazine - August 2022 - 30

The decentralized accountability enables all miners
to verify the quality of up-loaded models that are
recorded on the blockchain.
volume can be proportional to the size of the training data.
It
is noted that
the reward mechanism can be further
amended by combining consideration of the data size and
the quality of data samples. In this case, clients are responsible
for verifying the trustworthiness of local updates after
aggregation to address the situation that untruthful UEs
may exaggerate their sample sizes with abnormal local
model updates. Specifically, when clients calculate the
rewards for each training node, they can give scores/reputations
to the training nodes based on the model qualities. In
the next aggregation, nodes with low scores will be given
less weight, and they will be identified and gradually
ignored during learning. In practice, this can be guaranteed
by Intel's software guard extensions, allowing applications
to be operated within a protected environment, which has
already been used in blockchain technologies [14]. In addition,
miners can also obtain rewards from mining and
aggregating models, which can be treated as a gas tax in the
traditional blockchain.
A task publisher first broadcasts an FL task through an SC
to all of the clients. Consider that N clients dynamically bid for
this task and use PoW as the consensus mechanism in the
ALGORITHM 1 BLADE-FL algorithm.
Data: Number of communication rounds T, initial model
w0
, and proximal term μ in local learning.
1 Task publishing and client selection.
2 Initialization: t = 1 and wwi
=
3 while tT# do
4
5
6
7 ww ww() ||=+ -
Model broadcasting and aggregating:
Update aggregated model w()t
i
t
8
9
N
10 wwp
t
i
11
12
13
14
15
= / N .
i=1
i
t
Block mining and verification:
Each client starts mining its block that includes the
aggregated model and verifies the generated block.
Global model downloading:
Each client downloads the aggregated model from the
verified block and updates.
Reward allocation for learning and mining.
16 tt 1! +
Result: w
N
(T)
30 IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE | AUGUST 2022
argmin Fii
w `
i
n
2
as
i
Local training:
while i {1,2,, }Nf!
do
Update local model w()
t
i
as
t 1
-
||
2 j.
0 6 i
,
verification process. Thus, the pseudo-code of
the proposed BLADE-FL framework is outlined
in Algorithm 1.
III. Unique Issues and
Potential Solutions
In this section, three critical issues that the
proposed framework may be confronted with, namely, privacy,
resource allocation, and lazy clients, are described.
A. Privacy
In BLADE-FL, the roles of each client include mining and
training. To aggregate the global model, the trained local model
will be published among clients, which raises privacy issues.
Previous works [5]-[11] usually artificially assign the training
and mining tasks to two disjoint sets of clients, and they widely
adopt that the miners are always trustful. However, if an eavesdropper
exists in the wireless environment, the published information
of local models can cause privacy leakage. To address
this, a differentially private mechanism can be implemented at
the client side. In detail, the key steps are as follows.
❏ Each client sets up a self-required privacy level for itself
before training. For example, the ith client may have a local
privacy budget ie . Note that a small value of ie represents a
high local privacy level, and it will induce more additive
noises on the parameters.
❏ To achieve local differential privacy (LDP), each client will
add a random noise that follows a certain distribution on the
uploaded models. For example, a random Gaussian noise
N (, )0 2
v or a Laplace noise La ()p m will be added. Note that
a large noise power implies a high privacy level.
❏ Upon receiving the perturbed models, all of the clients can
aggregate the global model locally and store it in the generated
block. Because of the injected noise, the learning convergence
as well as the system performance will be
negatively affected. A trade-off between the privacy requirement
and learning performance needs further investigation.
In addition, a non-uniform allocation of additive noise over
communication rounds may improve learning performance;
for example, a decay rate for the noise power can be applied
when the learning accuracy between two adjacent communication
rounds stops improving [15].
B. Computing Resource Allocation
Since the computation resource is limited at each client, each
participant must appropriately allocate the resources for local
training and mining to complete the task. Specifically, more
computing resources can be devoted either to accelerate model
updating or block generation. To meet the specific task requirements,
such as learning difficulty, accuracy, and delay, each node
optimizes its allocation strategy to maximize its reward under
constraints of local capability.
According to the constraints, the computing resource allocation
can be formulated as an optimization problem under the
accurate mathematical model, as follows, in detail.
IEEE Computational Intelligence Magazine - August 2022

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