Membership inference attacks on synthetic mobility data
Membership inference attacks (MIAs) are a class of attacks that seek to extract sensitive information from a model's training dataset by analyzing the model's outputs and potentially revealing real records used during training. The scope of this thesis is to study, design, and implement a...
| Autor: | |
|---|---|
| Tipo de recurso: | tesis de maestría |
| Fecha de publicación: | 2024 |
| País: | España |
| Institución: | Universitat Politècnica de Catalunya (UPC) |
| Repositorio: | UPCommons. Portal del coneixement obert de la UPC |
| Idioma: | inglés |
| OAI Identifier: | oai:upcommons.upc.edu:2117/424756 |
| Acceso en línea: | https://hdl.handle.net/2117/424756 |
| Access Level: | acceso abierto |
| Palabra clave: | Machine learning Data protection Computer security Artificial intelligence Membership Inference Attacks Differential Privacy Shadow Modeling GAN Aprenentatge automàtic Protecció de dades Seguretat informàtica Intel·ligència artificial Àrees temàtiques de la UPC::Enginyeria de la telecomunicació::Telemàtica i xarxes d'ordinadors::Serveis telemàtics i de comunicació multimèdia |
| Sumario: | Membership inference attacks (MIAs) are a class of attacks that seek to extract sensitive information from a model's training dataset by analyzing the model's outputs and potentially revealing real records used during training. The scope of this thesis is to study, design, and implement a membership inference attack on a trajectory generator model. This includes a comprehensive review of the state of the art, describing the most prominent approaches related to MIA and mathematically formalizing the problem statement. We then provide two attack architectures under different assumptions by adapting and combining accepted solutions for trajectory data. Next, we evaluate the proposed architectures based on how well they infer real records and how they respond when the target model protects the training records with differential privacy mechanisms. Finally, we present a set of conclusions and potential directions for future research. |
|---|