• blockchain
• rewarding
• compensating contributions
• contribution-based aggregation method
• erc20
• token economy
• federated learning

Federated Learning allows multiple participants to collaboratively train a shared prediction model without exposing private datasets. However, it is vulnerable to adversarial attacks led by malicious users that aim to decrease the performance of the global model. Moreover, users are not encouraged to participate in collaborative learning because of the absence of incentives.
This thesis presents TOBE-LearneD, a decentralized framework that implements a mechanism to encourage users to collaborate and solve proposed federated learning tasks. We studied a contribution-based aggregation method to avoid poisoning attacks from malicious users. Exploiting public Blockchain technology, we implement a token-based economy to assign rewards weighted on participants' contributions, compensating the computational cost needed to train local models. Our framework allows users to propose a customizable Federated Learning task utilizing Smart Contracts, ensuring transparency and tamper-proof properties, thus preventing denial of reward to the participants from malicious manufacturers. In addition, we evaluate a prototype of TOBE-LearneD from a performance, economic and security perspective in a real-world use case related to the SmartGrid field.