IEEE Computational Intelligence Magazine - August 2022 - 27

this framework, the model aggregation process is fully decentralized and the tasks of training for FL and mining for
blockchain are integrated into each participant. Privacy and resource-allocation issues are further investigated in the proposed
framework, and a critical and unique issue inherent in the proposed framework is disclosed. In particular, a lazy client
can simply duplicate models shared by other clients to reap benefits without contributing its resources to FL. To
address these issues, analytical and experimental results are provided to shed light on possible solutions, i.e., adding noise
to achieve local differential privacy and using pseudo-noise (PN) sequences as watermarks to detect lazy clients.
uture wireless networks are expected to require very low
latencies and high reliability. Migrating machine learning
(ML) to end-user equipments (UEs) promotes these
requirements, giving them the capability of making decisions
based on locally acquired data, even if it loses connectivity to
the network. Since data available at a given end-user device is
typically limited, the training of on-device ML models can benefit
from data exchange among UEs [1].
However, directly exchanging data among UEs presents
F
risks of privacy leakage and information hijacking [2]. To
reduce this risk, federated learning (FL) has been proposed,
which is an ML framework that trains an artificial intelligence
(AI) model across multiple UEs holding local datasets. In particular,
distributed UEs train ML models locally, sharing their
model parameters with a central server where these local models
are aggregated into a global model. In this way, FL allows
UEs to cooperatively learn a global model without exchanging
their data directly. FL has been applied in practical settings,
including health care and autonomous driving [3].
Although FL offers advantages in latency and privacy
enhancement, it suffers from several limitations. First, in the FL
process, it is assumed that the aggregator is trustworthy and will
make fair decisions in terms of user selection and aggregation.
However, this assumption is not always satisfied in practical situations
where a biased aggregator can intentionally favor a few
selected UEs, thereby biasing learning performance [1]. Second,
although the aggregator has access only to the models trained by
its UEs, private client data can still be inferred from those models.
Thus, if the aggregator is compromised, privacy leakage happens.
Lastly, the conventional FL architecture is vulnerable to malicious
clients that can attack learning via model poisoning [4].
As a secure technology, blockchain has the capability to tolerate
a single point of failure with distributed consensus, and it can
further implement incentive mechanisms to encourage participants
to effectively contribute to the system [5]. For these reasons,
blockchain has been introduced into FL to mitigate its
aforementioned limitations. For example, [5] introduced a blockchained
FL architecture to verify uploaded model parameters
and investigated related system performance indices, such as
learning delay and block generation rate. Moreover, [6] proposed
a privacy-aware architecture that uses blockchain to enhance
security when parameters of ML models are shared among UEs.
In addition, the authors of [7] proposed a high-level framework
by enabling encryption during model transmission, and [8] further
applied this framework in a military setting. With the
advanced features of blockchain, such as tamper-resistance, anonymity,
and traceability, an immutable audit trail of ML models
can be created for greater trustworthiness in tracking and proving
provenance [9]. In addition, security and privacy issues arising
in the decentralized FL framework are investigated in [6], [10],
[11], which delegated the responsibility of storing ML models to
a trust community in the blockchain. However, the assumption
of a trust community may incur the same privacy issues when
ML models are transmitted over the air, and the credibility of this
community also needs further verification. In addition, these
works either have not completely clarified and fully addressed
incidental issues, such as the long learning delay and impact of
blockchain forking on FL, or are difficult to apply.
In the present work, a blockchain-assisted decentralized FL
(BLADE-FL) framework, which can overcome the single point of
failure problem, is proposed in detail. In addition, several residual
issues that exist in the BLADE-FL framework are further investigated,
and related solutions are provided. The rest of this paper is
organized as follows. The design of the BLADE-FL framework is
presented in Sec. II, and residual issues, including privacy, resource
allocation, and lazy clients, are investigated in Sec. III. In Sec. IV,
extensive experimental results are provided to show the effectiveness
of the corresponding solutions. Finally, promising future directions
are suggested and conclusions are drawn in Sec. V.
II. BLADE-FL Framework
With the aid of blockchain, the aim is to build up a secure and
reliable FL framework. To ensure this, the model updating process
of FL is decentralized at each participating client, which is
robust against the malfunction of traditional aggregators. In this
section, the BLADE-FL framework, as well as how it achieves
dynamic client selection and a decentralized learning aggregation
process is presented.
The BLADE-FL framework is composed of three layers. In
the network layer, the network features a decentralized peerto-peer
(P2P) network that consists of task publishers and
training clients, wherein a learning mission is first published by
a task publisher and then completed by the cooperation of
several training clients. Different from previous work, in which
model aggregation occurs in a trust community in the blockchain
[5]-[11], a fully decentralized framework is realized in
which each client must train ML models and mine blocks for
publishing aggregating results. In the blockchain layer, each
FL-related event, such as publishing a task, broadcasting learning
models, and aggregating learning results, is tracked by
blockchain. In the application layer, the smart contract (SC)
and FL are utilized to execute the FL-related events. Next, the
workflow and key components of the BLADE-FL framework
are presented.
AUGUST 2022 | IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE 27

IEEE Computational Intelligence Magazine - August 2022

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