Personalized Federated Learning with Liquid Neural Networks

Master's Thesis

Description
Federated Learning (FL) enables model training across multiple devices/sites without centralizing raw data—highly relevant for audio/speech, where data can be sensitive. FedCfC combines FL with CfC models and targets personalized models that improve the capture of local characteristics (e.g., microphone response, environment, speaker/machine variability). This thesis evaluates FedCfC in a controlled FL simulation: how personalization impacts performance, how stable training is under non-IID client data, and how FedCfC compares to standard FL approaches (e.g., FedAvg + fine-tuning). The downstream task is chosen to remain audio or speech-related (e.g., phoneme prediction; optionally audio forecasting or reconstruction-based anomaly detection).

Task
Implement or integrate FedCfC into an FL simulation setup (e.g., using PyTorch Lightning), define an audio- or speech-oriented multi-client scenario (non-IID partitions), and run systematic experiments (baselines, ablations, personalization strength). Evaluate task performance, convergence, communication/compute cost, and robustness.

Prerequisites

• Strong Python and PyTorch skills
• Fundamentals of audio/signal processing (STFT, mel-spectrograms, feature design)
• Interest in deep learning
• Basic understanding of federated learning