Name: JOÃO PAULO DE BRITO GONÇALVES
Publication date: 26/03/2025
Examining board:
Name |
Role |
|---|---|
| ANTÔNIO AUGUSTO ARAGÃO ROCHA | Examinador Externo |
| JOHANN MARQUEZ-BARJA | Coorientador |
| JOSÉ MARCOS SILVA NOGUEIRA | Examinador Externo |
| RAFAEL PASQUINI | Examinador Externo |
| RODOLFO DA SILVA VILLACA | Presidente |
Pages
Summary: Service Level Agreements (SLAs) are formal contracts between consumers and service
providers that describe the parameters of Quality of Service to be provided. In addition to
the nature of the service itself and its expected level of performance, an SLA also specifies
procedures for monitoring and reporting issues, time limits for resolving issues, and the
consequences for both customers and service providers when clauses are violated. Edge
computing brings a new paradigm in which computing, storage, and bandwidth resources
are shared as close as possible to mobile devices or sensors, generating vast amounts of
data. A parallel trend is the rise of smartphones and tablets as primary computing devices
for many people. The powerful sensors in these devices (including cameras, microphones,
and GPS), combined with the fact that they are mobile, mean that they have access to
unprecedented amounts of private and diverse data, and the sensitive nature of the data
means that there are risks and liabilities in storing it in a centralized location, which
makes it difficult to verify SLA compliance using this private data. Furthermore, several
countries now require technology companies to handle user data with care, in accordance
with user privacy laws. The European Union’s GDPR is a prime example. Thus, to address
the data privacy requirements for some data on these devices, we propose the use of FL so
that specific data about services accessed by clients does not leave the source machines.
FL is a new subfield of ML that allows training models without collecting the data itself.
Instead of sharing data, users train a model collaboratively by sending only weight updates
to a server. However, naively using FL in the above scenarios exposes it to a risk of
corruption, intentional or unintentional, during the training phase. This is due to the lack
of monitoring of the training and the difficulty in verifying the quality of the training
datasets. To improve the security of FL systems, we propose a blockchain-based framework
in an edge computing scenario. Blockchain, with its immutability and traceability, can
be an effective tool to prevent malicious attacks in FL. More specifically, the immediate
updates made by each participant to their local model can be chained into the distributed
ledger provided by a blockchain, so that these model updates are audited and malicious
trainers can be removed from the system. We also apply blockchain to build a reward
mechanism in FL to enable an incentive strategy for trainers. We validate our approach by
demonstrating our solution to protect sensitive data by deploying a Proof of Concept (PoC)
and evaluate performance metrics in different scenarios through extensive simulations.
