IEEE Consumer Electronics Magazine - November/December 2022 - 69

get control over the end devices and act
maliciously.
BACKGROUND AND RELATED WORK
Brief Overview of Federated Learning (FL)
FL is a distributed model where a general or
pretrained model is distributed to clients initially.
With the local data, all the clients personalize
their model by performing ML tasks locally
and send its updated parameters to the FL
server. The server then aggregates all the
updates received from the clients and performs
ML tasks and finally distributes the updated
global model to the clients7. The process of
updating global model is iterative. The parameters
at each client are optimized by minimizing
the local objective function such as stochastic
gradient descent (SGD). The optimized parameters
from client devices are sent to the server
and clients receive the updated global model
(e.g.,4).
Brief Overview of Blockchain (BC)
Although the concept of Blockchain has been
in existence since 2008, it has been in the limelight
just recently, thanks to the widespread success
of cryptocurrency platforms (blockchain
applications like Bitcoin and Etherium8,9). BC
technology is based on distributed digital ledge
that uses public key cryptography and consensus
algorithms as core components to offer
unparalleled benefits of security, transparency,
and traceability. In blockchain, each block holds
multiple records of transactions that occurred
within some time period and a block header that
contains two important information, which are
the previous block address and Merkle root8.
Merkle tree root is a kind of fingerprint of all the
information on the BC that is useful to validate
the consistency of all the information of the BC.
It changes even if any tiny bit of information on
the BC changes. Merkle root is basically a binary
tree root created from combining hashes of all
the transactions.
Overview of FL and BC and Related Work
FL seems an effective solution regarding preserving
privacy4-6, however, this approach is
still vulnerable to poisoning and reverse
November/December 2022
engineering attacks. An adversary can reveal the
privacy of a user by spoofing on model updates
sent by the client. Moreover, if the adversary
gains control on the aggregating server, he/she
can get comprehensive knowledge of the history
of update parameters of devices and the structure
of the global model. With these information,
adversaries can reveal the privacy of users
through reverse engineering. Further, with the
full control of the central server, adversaries
might forge the global model parameters, which
in turn might affect the local model parameters
of the end devices. In effect, aggregating such
local updates might degrade the quality of the
global model significantly. Even if adversaries do
not have control over the end device or server,
the model parameters might still be forged while
in transit between client and server. On the
other hand, due to vulnerabilities of IoT sensors
and the wireless communication standard they
use, devices are prone to be attacked by adversaries.
If the adversary gains control over the
devices and forges the local training data or
model parameters, the accuracy of the global
model is affected. Some existing works in the literature
have addressed some of these security
and privacy issues using differential privacy10,
secure multiparty computation11 and blockchain
technology for IoT devices12. However none of
these methods consider blockchain empowered
trustless, and secure FL framework for IoV where
network topology changes frequently, which is
proposed in this article. Specifically, the proposed
BC enabled FL framework consists of registered
vehicles, RSUs, cloud service providers
(CSPs), and authorities. These participants
(except vehicles) perform roles of miners, verifiers,
light nodes based on their level of computing
resources whereas vehicles do the role of
network users only. The updated model parameters
(including aggregated parameters) are
broadcast through the BC network and the integrity
of parameters are verified by verifiers, which
are then bundled into a valid block by a miner.
Each block stores model parameters belong to a
particular communication round of a particular
learning task. Storing the history of exchanged
parameters in BC ensures robustness, however,
it also incurs significant storage and makes the
whole system inefficient. To address this, we
69
IEEE Consumer Electronics Magazine - November/December 2022

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