A Novel Privacy-Preserving Method for Graphs with Autoencoder Representations

• Ano de defesa: 2025
• Autor(a) principal: Oliveira, Gustavo Lima de
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Aprendizado de preservação de privacidade; Autoencoder para grafos; Graph autoencoder; Graph-based representation; Privacy-preserving learning; Representação em grafos
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/55/55134/tde-13082025-141242/

Sharing data publicly or among entities with similar interests can lead to the development of more robust applications and the resolution of complex problems. In the context of machine learning, for instance, training models with diverse data sources increases the efficiency and effectiveness of the solutions developed. With the massive use and generation of data in computational systems, it is becoming increasingly essential to implement solutions that enable the analysis and extraction of latent information from large volumes of data. However, a growing concern in this scenario is the privacy and security of users sensitive information. Around the world, laws and regulatory tools have emerged to prevent leaks and protect the integrity of stored information. In addition to leaks resulting from social engineering strategies, data that have been anonymized using traditional methods can still be exposed. With the evolution of machine learning models, data attacks have also become more sophisticated. In this context, graph-based strategies have shown promise, as they implicitly capture the attributes and relationships between instances in the dataset. In light of the challenges presented, this dissertation proposes an advancement in privacy preservation within graph-based strategies. Through the use of autoencoders, we aim to penalize the correct classification of sensitive attributes while maintaining the utility of the data for subsequent tasks. The method has been applied in various contexts and data sources, such as medical information and social media, and has demonstrated significant results compared to traditional approaches, especially when there is a strong correlation between useful and private data. Finally, among the main contributions of this dissertation are: (i) a method for generating k-NN graphs that takes privacy into account; (ii) a specific method for medical data that utilizes a heterogeneous information network; and (iii) an approach aimed at unstructured data from social networks.