Advancing Secure Data Management through the Model for On-Air Split Federated Learning On Data Silos

The challenge of securely managing sensitive information on resource-constrained devices is markedly evident in critical applications such as drone swarming in defence scenarios. Existing frameworks lack efficiency and confidentiality, particularly when devices operate in isolation, devoid of key parameter sharing for enhanced data processing and model learning. This limitation curtails performance and restricts the wider applicability of current models. A Model for On-Air Split Federated Learning has been developed to address these challenges, featuring an On-Air Server Collector. This model allows for the retention of data on resource-constrained devices while offloading requisite computations into mutually exclusive data silos, thereby establishing a robust data management framework. It significantly improves operational efficiency and ensures data integrity across a spectrum of domains. The architecture facilitates parallel client computation in environments where each client’s data is distinct and where sharing raw data is stringently restricted. It yields outcomes comparable to those obtained with a centralized Deep Neural Network (DNN) model. The model proves advantageous in a range of applications, including decentralized training for drone swarming, biometric voice recognition, precision medicine, and consumer service personalization, where devices are resource-constrained and contain sensitive information necessitating restricted sharing. This model offers a compelling value proposition for companies aiming to undertake decentralized training of machine learning models in settings where hardware limitations and privacy concerns constrain the sharing of training data and data labels.