Aprendizado federado e mecanismos de privacidade diferencial para preservar dados sensíveis em treinamento de modelos de machine learning

Abstract

As mobile devices continue to evolve, their sensing and processing capabilities have reached unprecedented levels of sophistication. Coupled with advances in deep learning, this progress has opened up diverse opportunities for high-impact applications, particularly in areas such as healthcare and automotive systems. The traditional approach to cloud-based machine learning (ML) requires the aggregation of data on a centralised server or data centre. However, this approach raises critical issues of latency, inefficient commu nication and data privacy. The prevailing technologies for ML in mobile and distributed networks still require the disclosure of personal data to external parties. Recently, the concept of Federated Learning has been introduced in the face of increasingly stringent data protection legislation and growing privacy concerns. In FL, end devices use their local data to train an ML model that is then required by the server. The devices then send updates to the model, rather than the raw data, to the server for aggregation. Federated Learning (FL) has the potential to be a revolutionary technology in mobile networks as it enables collaborative training of an ML model while preserving client privacy. However, private information can still be discovered by analysing parameters loaded from clients, for example, weights trained on deep neural networks. FL is susceptible to inference attacks, which can originate from any party that contributes to the training process. In this sense, Differential Privacy (DP) is a technique that introduces controlled statistical noise into the data, preventing the inference of participant-specific information, even in adverse circumstances. In this work, DP was used in combination with different noise mechanisms, such as Laplacian and Gaussian, to analyse how DP affects model accuracy while trying to mitigate these attacks during training in a Federated Learning environment. This integration aims to achieve a balance between privacy and model performance.