IEEE Consumer Electronics Magazine - November/December 2022 - 63

Federated Training
Given a TFF task, the cloud checks the willingness
of vehicular participants to be involved in
the training process and rejects participants
with possible failures. Then the cloud initializes
the parameters of the global model and broadcasts
the global model' gradient to the selected
local participants to start the local training using
their private data.17 Once the local training
is completed, the local parameter should be
uploaded to the cloud. However, the local
parameters contain information that enables the
attackers to infer the local data. Hence, preserving
the privacy of local parameters is essential
to guarantee the Fed-GCRN framework. Previous
studies employ MPC and HE techniques to maintain
the privacy of the FL scheme, yet these techniques
are requiring heavy computation and
communication which makes them inadequate
to cope with the resource-constrained nature of
IoVT applications.
The ð2; dÞ-DP mechanisms have been introduced
to solve this issue, yet it suffers froma great
tradeoff between privacy and efficiency.18 To
tackle this, the Fed-GCRN introduces an adaptive
DP (ADP) mechanism along with the random data
subsampling method. In particular. The ADP is
designed by combining the randomized DP,
Gaussian DP, and Laplace DP methods to act like
the typical building elements for preserving privacy,
where the composition theorem19 is
employed to calculate the values of 2 and d.
Besides, rather than probing all the local traffic
datasets, an arbitrary subsample can be inquired.
In this way, when an ð2; dÞ-P is employed to probe
random subsample from a traffic data of n instances,
then an enhanced ð20; d0Þ factors could be
given based on the type of random subsampling.
When the size m of the subsample is less than n,
then value of ð20; d0Þ is more improved than the
original ð2; dÞ in such way that it is tinier and
hence is the devoted privacy budget, wherein the
injected noise is deemed as originated from different
sources. This implies adding a small amount
of noise and a smaller privacy budget expenses at
the cost of reading a little subsample of the data.
Once the ADP is applied to the local parameters,
the vehicular participants submit them to the
cloud. the cloud aggregates the uploaded noised
local parameters to upgrade the global model. the
cloud use FedAVG algorithm to lessen the overheads
of uplink communication. Once the global
model updated, the new parameters are saved
and broadcasted back to the participants. This
process is interchangeably repeated till the FedGCRN
converges.
EXPERIMENTS AND ANALYSIS
Experimental Design
The simulation experiments are coded
using Python 3.7 with a Dell workstation armed
with a CPU of Intel (R) Xeon (R) CPU E5-2670
0@ 2.60GHz, a memory of 256 GB, and a graphical
processing unit of NVIDIA Quadro K2200.
Which is running using 64-bit windows 10. The
Fed-GCRN is evaluated using traffic flow data of
the first two months of 2018 from the public
dataset aggregated from the Caltrans performance
measurement system20 with over 39,000
different gauges spanning the freeway system
throughout the majority of metropolitan
regions in the State of California. The data is
splitted into 80% for training and 20% for testing.
Moreover, the root mean square error,
mean absolute error, and mean absolute percentage
error are employed as metrics for evaluating
IEEE Consumer Electronics Magazine - November/December 2022

Table of Contents for the Digital Edition of IEEE Consumer Electronics Magazine - November/December 2022
