Important assembly of researchers and practitioners in the field of machine learning and artificial intelligence, the NeurIPS-2020 Workshop on Scalability, Privacy, and Security in Federated Learning With an eye on scalability, privacy, and security especially, it sought to investigate the possibilities and difficulties in the newly developing discipline of Federated Learning (FL).

The main ideas and workshop debates are broken out here

1. Federal Learning (FL) Introduction

Federated Learning is a distributed machine learning technique whereby several participants—such as mobile devices, edge nodes, or businesses—can cooperatively train a machine learning model while maintaining the training data localised. Since data does not have to leave its original place, this method answers issues about data privacy and bandwidth restrictions. Rather, participants share just model updates.

2. Challenges of Scalability

Scalability presents one of FL's toughest obstacles. Hundreds or thousands of devices join in the training process in large-scale federated systems. Important problems that were underlined consist

Efficiency of Communication: The communication overhead of exchanging model updates rises dramatically as the participant count rises. We spoke about methods for lowering communication costs including asynchronous updates, model compression, and aggregation strategies.

Devices in a federated learning network could differ in processing power, network bandwidth, and battery life. This variability might hamper model training, hence it is crucial to build algorithms resistant to variations in device capability.

Training a global model can be challenging since each device generates often non-iid (independent and identically dispersed) data. We investigated solutions for non-iid data including tailored federated learning methods.

3. Federal Learning Privacy Issues

Since Federated Learning seeks to teach models without sharing raw data, privacy is fundamental in the approach. Still, privacy carries certain hazards: Data Leakage: Particularly in iterative training systems, model updates could cause data leakage even when data stays on the device. Methods to safeguard user privacy were underlined include secure aggregation and differential privacy.

Attacks of model inversion and membership inference can expose private information about the data utilised for model training. Researchers spoke on defensive tactics including adversarial training and other privacy-preserving devices to stop these attacks.

Protocolues for maintaining privacy: Presenting cryptographic techniques such as Homomorphic Encryption and Secure Multi-Party Computation (SMPC) as means to guarantee that model updates are computed without disclosing sensitive data to any party.