IEEE Consumer Electronics Magazine - March/April 2023 - 32

Special Section on Security, Trust and Services for AI-Empowered Sensing in the Next Generation IoT
article, we utilize the stochastic gradient
descent algorithm to find the model parameters
that correspond to the most optimal point
between actual outputs and predicted outputs.
6) Local training: Upon receiving the energy
demand prediction global model, CSs first
use their own ski to decrypt the global model
and start to train energy demand prediction
local models by using EV charging data.
7) Model upload:At this stage, each CS uploads
its trained local model updates to the CSP.
FIGURE 3. Mutual authenticationbetweenCSandCSP.
equal to
QpubÞghi
PN
i¼1 H2 mi;e Si;ð
ð
H1ðCSIDiÞQþ
Þ, where N is the number of CSs and
i means the order of the CS. Once all the CSs
are verified, the CSP sends a signature SigCSP
to the respective CSs for identity proof. For
SigCSP generation, the CSP first chooses two
random positive integer mCSP;rCSP and
calculates grCSP
;H2ðmCSP;RCSPÞ; and ðrCSP þ
hCSPÞSCSPID corresponding to RCSP;hCSP;
and SCSP, respectively. Subsequently, CSP
uses its private key SCSPID to restore corresponding
ski used for the following global
model encryption. Finally, the CSP returns its
signature SigCSP, which is made up of
ðhCSP;SCSP;tCSPÞ, where tCSP is the timestamp.
On the side of CSs, the verification
process is similar with CSP. CSs need check
the timestamp of CSP and the validness
of hCSP, which is H2 mCSP;e SCSP;H1
ðCSPIDÞQþQpubÞghCSP
ð
ð
Þ. The detailed steps
of mutual authentication are presented in
Figure 3. After the first training turn, each CS
will be distributed a trust value calculated by
the size and accuracy of upload model. Note
that when the trust value is lower than 60%,
corresponding CS will be considered out of
order and dequeued.
5) Task initialization: After successful mutual
authentication, the CSP distributes an initial
energy demand prediction global model trained
by some historical charging data to the
CSs. Thismodel is encrypted by the ski and the
encryption algorithm is AES. Note that in this
32
8) Model evaluation:To accelerate the training
speed and defend against some potential
attacks, in this step, the CSP needs to evaluate
the size and accuracy of uploaded models
to calculate the latest trust value and
issue award for the corresponding CSs.
1) Trust value calculation: The trust value
(denoted as tv) depends on the ith model
size Di and accuracy li, which can be
defined as ð1 liÞDi=ðDi þDi liÞ. Note
that only the CSs who own trust value
over 60% can join the next training turn.
2) Award issue: To encourage CSs' motivation,
some award in the form of electricity
will be repaid to the CSs according to their
contribution portion. The related award
calculation formula can be defined as tv
1000. Note that the unit of award is kWh.
9) Model issue:When the model evaluation is finished,
unqualified models will be removed.
ThenCSPbegins to aggregate all models to generate
a more accurate global energy demand
prediction model, then reissued to the CSs.
Theabovementioned steps 6-9will berepeated
until the optimal energy prediction model
is found by CSP. Furthermore, this final highprecisionmodel
will help CSP to generate a reasonable
energymanagement strategy.
Security Analysis and Performance
Evaluation
Security Analysis
In this section, we analyze the reliability of our
proposed systemaccording to the possible attacks
discussed in Section " Existing Threats in FL " .
1) Information exploiting attacks defense: In our
proposed model, only the scrutinized CSs
can participate in our energy prediction
IEEE Consumer Electronics Magazine

IEEE Consumer Electronics Magazine - March/April 2023

Table of Contents for the Digital Edition of IEEE Consumer Electronics Magazine - March/April 2023