Defesa de Tese de Doutorado: Private and Secure Federated Learning Protocol Based on Asymmetric DC-Nets

Orientadores:
Fábio Borges de Oliveira - Laboratório Nacional de Computação Científica - LNCC

Banca Examinadora:
Fábio Borges de Oliveira - Laboratório Nacional de Computação Científica - LNCC (presidente)
Allan Jonathan da Silva - Laboratório Nacional de Computação Científica - LNCC
Erick Giovani Sperandio Nascimento - University of Surrey
Raphael Carlos Santos Machado - Instituto Nacional de Metrologia, Qualidade e Tecnologia - INMETRO

Suplentes:
Renato Portugal - Laboratório Nacional de Computação Científica - LNCC
Lisandro Zambenedetti Granville - Universidade Federal do Rio Grande do Sul - UFRGS

Resumo:

Federated Learning (FL) enables collaborative machine learning across multiple client s without centralizing raw data, offering a promising approach for privacy-sensitive domains. However, standard FL protocols are vulnerable to inference attacks. While existing Secure Aggregation (SA) methods address this, they typically suffer from quadratic O(N²) communication overhead or the prohibitive computational costs of Homomorphic Encryption (HE). To overcome these fundamental bottlenecks, this thesis proposes and evaluates ADC-Fed, a highly scalable SA protocol that synergizes Asymmetric DC-Nets (ADC-Nets) with Multi-Secret Sharing (MSS) and Vandermonde interpolation. By completely abandoning heavy homomorphic operations, which were empirically demonstrated to impose a strict computational wall in an evaluated HE-based baseline (HE-ADC), ADC-Fed achieves information-theoretic privacy with strictly linear O(N) cryptographic communication complexity. Within the reported simulation environments and benchmark datasets, the evaluation demonstrates that ADC-Fed provides exact r econstruction of the intended aggregate, preserving the exact predictive utility of the non-private baseline without requiring interactive peer-to-peer dropout recovery. These results hold under the stated honest-but-curious server threat model, subject to the Minimum Honest Client Bound (MHC) where privacy is maintained provided at least two clients remain honest. Ultimately, this research validates ADC-Fed as a robust, high-performance architecture for privacy-preserving Federated Learning in resource-constrained and volatile network environments.