IEEE Computational Intelligence Magazine - August 2022 - 28

A. Workflow
As shown in Fig. 1, the workflow of the proposed framework
consists of the following steps.
❏ Step 1: Task publishing and node selection. A task publisher
broadcasts an FL task by deploying an SC over the blockchain
network. In the deployed SC, the task publisher must
deposit a reward as a financial incentive for the learning
task. The SC selects available training nodes to participate in
this learning task.
❏ Step 2: Local model broadcast. Each training client runs its
local training by using its own data samples, and it broadcasts
local updates and the corresponding processing information
(e.g., computation time and local data size) over the P2P
network. Privacy leakage may happen during this transmission,
and this issue is further investigated in Sec. III-A.
❏ Step 3: Model aggregation. Upon receiving the local
updates from other training nodes before a pre-set timestamp,
each client updates the global model according to the
aggregating rule defined in the SC.
❏ Step 4: Block generation. Each training client changes roles
from trainer to miner and begins mining until it either finds
the required nonce or receives a generated block from other
miners. The learning results are stored in the block as well.
When one miner generates a new block, other clients verify
the contents of this block (e.g., the nonce, state changed by
the SC, transactions, and aggregated model). The resourceallocation
issue happens in each client in this step, and related
discussions will be given in Sec. III-B.
❏ Step 5: Block propagation. If a block is verified by the majority
of clients, this block will be added on the blockchain
and accepted by the entire network. The lazy client issue
occurs in this step and is further investigated in Sec. III-C.
❏ Step 6: Global model download and update. Each training
client downloads the aggregated model from the block and
performs updates before the next round of learning.
❏ Step 7: Reward allocation. The SC deployed by the task
publisher rewards the training clients according to their
contributions to the learning task.
Block
Verification
Smart Contract
Design
New Block Accountant
Book
Mining Process
Step 4: Block
Generation
Step 7: Providing
Rewards
Issues:
Step 3: Model
Aggregation
Model Pool
Mining
Task
Publisher
Role Changing
Step 2: Local
Model Broadcast
Step 1: Task
Publishing and Client
Selection
Learning
Client
Federated Learning Process
Local
Learning
Global Model
Update
Step 6: Global
Model Download
Model Pool
Mining
Computing Resource Allocation:
For each client, a reasonable
resource allocation for learning and
mining should be further planned.
Privacy Leakage: Upon receiving
the models broadcast from other
clients, it is possible that a client
can infer the private information of
other clients.
Lazy Clients: A lazy client can
simply duplicate models shared
by other honest clients in order to
reap benefits without contributing
its computing resources.
Step 5: Block
Propagation
P2P Network Layer
Blockchain Layer
Application Layer
FIGURE 1 BLADE-FL workflow.
28 IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE | AUGUST 2022
IEEE Computational Intelligence Magazine - August 2022

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