Ferreira, Bernardo Luís da SilvaÂngelo, Daniel José Ferreira2023-03-282023-03-2820232022http://hdl.handle.net/10451/56882Tese de Mestrado, Engenharia Informática, 2022, Universidade de Lisboa, Faculdade de CiênciasTor is one of the most popular anonymity networks in the world. Users of this platform range from dissidents to cybercriminals or even ordinary citizens concerned with their privacy. It is based on advanced security mechanisms that provide strong guarantees against traffic correlation attacks that can deanonymize its users and services. Torpedo is the first known traffic correlation attack on Tor that aims at deanonymizing onion services’ (OS) sessions. In a federated way, servers belonging to ISPs around the globe can process deanonymization queries of specific IPs. With the abstraction of an interface, these queries can be submitted by an operator to deanonymize OSes and clients. Initial results showed that this attack was able to identify the IP addresses of OS sessions with high confidence (no false positives). However, Torpedo required ISPs to share sensitive network traffic of their clients between each other. Thus, in this work, we seek to complement the previously developed research with the introduction and study of privacy-preserving machine learning techniques, aiming to develop and assess a new attack vector on Tor that can preserve the privacy of the inputs of each party involved in a computation, allowing ISPs to encrypt their network traffic before correlation. In more detail, we leverage, test and assess a ML-oriented multi-party computation framework on top of Torpedo (TF Encrypted) and we also develop a preliminary extension for training the model with differential privacy using TF Privacy. Our evaluation concludes that the performance and precision of the system were not significantly affected by the execution of multi-party computation between ISPs, but the same was not true when we additionally introduced a pre-defined amount of random noise to the gradients by training the model with differential privacyengTorCorrelação de tráfegoPreservação de privacidade em aprendizagem automáticaComputação multipartidáriaPrivacidade diferencialTeses de mestrado - 2023master thesis203498747