Co-Utile protocols for decentralized computing

Identificador:  TDX:4278

Handle: https://hdl.handle.net/20.500.11797/TDX4278

Autores:  Manjón Paniagua, Jesús Alberto

Resumen:
Decentralized computing is a paradigm that distributes computing tasks and decision-making across a network, offering benefits like resilience, scalability, privacy, and reduced reliance on central authorities. It can address the tension between GDPR and big data analytics, as GDPR aims to protect the privacy rights of individuals and ensure the responsible handling of personal data while big data analytics requires collecting and processing large amounts of personal data to uncover correlations, trends, and predictive models. Decentralized protocols such as Federated Learning or Fully Decentralized Federated Learning offer a solution to this dilemma by removing the need to exchange data from client devices to global servers. Instead, the raw data on edge devices are used to train the models locally, increasing data privacy. This thesis focuses on one of the most important issues in this type of protocol: how to ensure that all agents involved perform as expected. Nodes that deliberately deviate may do so to attack the system or simply to take advantage of it without contributing. Based on the notion of co-utility, we have designed several ethics-by-design frameworks to solve the conflict between privacy and some security properties in three different decentralized and privacy-preserving computing scenarios: i) Federated learning; ii) Fully decentralized learning; and iii) Multi-party computation. These types of protocols facilitate collaboration between multiple parties or entities to achieve a common goal and operate under certain trust assumptions.