Licentiate Thesis Proposal-Samaneh

Abstract:

The widespread adoption of centralized machine learning (ML) techniques has revolutionized data analysis and decision-making across industries. Nevertheless, concerns regarding privacy have arisen due to the reliance of centralized ML on vast amounts of personal data and the General Data Protection Regulation (GDPR). To address these concerns and ensure compliance, federated learning (FL) has emerged as a promising solution. FL enables ML model training on user devices or edges, eliminating the need for data centralization and preserving sensitive information.

FL has gained attention in several emerging applications, such as autonomous vehicles, finance, healthcare, and entertainment, enabling collective model training while preserving privacy. In this study, we focus on Speech Emotion Recognition (SER) as a use case. SER has been utilized in diverse applications, from mental health and education to entertainment and customer service. Recognizing and understanding human emotions from speech opens up opportunities for improved experiences, personalized interventions, and better decision-making. In our study, FL is employed for SER to ensure privacy by keeping data on user devices, allowing speech data to remain local during training while transmitting only model parameters for aggregation.

However, FL introduces new privacy concerns when transmitting SER model parameters. Adversaries exploit various attacks, such as data reconstruction, property inference, and membership inference, aiming to infer private information from the shared model parameters. To address this issue, additional privacy mechanisms have been proposed alongside FL to safeguard such applications. While Differential Privacy (DP) is a commonly used mechanism in FL, applying DP to SER faces challenges, often reducing accuracy due to noise added to voice data. An alternative approach to ensure privacy and accuracy in SER applications is diverse homomorphic encryption methods, such as Paillier homomorphic encryption (PHE). PHE protects user privacy while maintaining SER model accuracy. However, employing PHE in FL for SER applications presents challenges in terms of performance impact such as increased communication traffic, computation time, and reduced accuracy. These challenges require careful consideration to ensure privacy and performance in FL setups for SER applications.

This research proposal aims to explore and address these challenges through two innovative methods. The first one is called LDP-FL with CSS, combining Local Differential Privacy (LDP) with a novel client selection strategy (CSS). By leveraging CSS, we aim to improve the representation of updates and mitigate the adverse effects of noise on SER accuracy while ensuring client privacy through LDP. Secondly, we introduce a novel approach called Secure and Efficient Federated Learning (SEFL) for SER applications. Our proposed method combines Paillier homomorphic encryption (PHE) with a novel gradient pruning technique. This approach enhances privacy and maintains confidentiality in FL setups for SER applications while reducing communication traffic and computation time without significantly sacrificing model accuracy.

To evaluate the effectiveness of our developed methods in real-world scenarios, we are designing a specific testbed specifically for evaluating SER in FL environments. The testbed comprises multiple devices, including edge devices with limited resources and laptops representing local clients. It also includes a central server for model aggregation and a communication network to simulate the transmission of model parameters.