Authors:
Aftab Akram
1
;
Clémentine Gritti
2
;
Mohd Hazali Mohamed Halip
3
;
4
;
Nur Diyana Kamarudin
3
;
4
;
Marini Mansor
3
;
Syarifah Bahiyah Rahayu
3
;
4
and
Melek Önen
1
Affiliations:
1
Department of Digital Security, EURECOM, 450 route des Chappes, 06410 Biot, France
;
2
CITI Lab, INSA Lyon – Inria, 69100 Villeurbanne, France
;
3
Faculty of Defence Science and Technology, University of Malaysia, 57000 Kem Sungai Besi, Kuala Lumpur, Malaysia
;
4
Cyber Security and Digital Industrial Revolution Centre, National Defence University of Malaysia, 57000 Kem Sungai Besi, Kuala Lumpur, Malaysia
Keyword(s):
Federated Learning, Byzantine Nodes, Secure Aggregation, Privacy, Robustness, Blockchain.
Abstract:
In Federated Learning (FL), clients collaboratively train a global model by updating it locally. Secure Aggregation (SA) techniques ensure that individual client updates remain protected, allowing only the global model to be revealed while keeping the individual updates private. These updates are usually protected through expensive cryptographic techniques such as homomorphic encryption or multi-party computation. We propose a new solution that leverages blockchain technology, specifically the Secret Network (SN), to provide privacy-preserving aggregation with aggregate integrity through Smart Contracts in Trusted Execution Environments (TEEs). Moreover, FL systems face the risk of Byzantine clients submitting poisoned updates, which can degrade the model performance. To counter this, we integrate three state-of-the-art robust aggregation techniques within the Smart Contract, namely Krum, Trim Mean and Median. Furthermore, we have evaluated the performance of our framework which rema
ins efficient in terms of computation and communication costs. We have also exhibited similar accuracy results compared to state-of-the art scheme named SABLE.
(More)