FedAqua

Exploring the potential for cross-industry collaboration in the Norwegian aquaculture sector using federated machine learning on AquaCloud data.

Federated Learning in Aquaculture

The aquaculture industry faces significant challenges that require innovative solutions, particularly when dealing with sensitive data and the need for collaboration between competitors. AquaCloud AS has played a key role in standardizing data collection and facilitating data sharing across the industry. By developing standards for sensor data, fish health data, and environmental data, AquaCloud ensures that information is collected and reported consistently.

Given the sensitivity of some of the data, it is crucial to utilize AquaCloud's resources safely. Federated machine learning offers a promising approach by enabling collaborative data analysis while maintaining data privacy and security. This secure sharing of models and learning from the unique AquaCloud centralized data platform, while data remains distributed, not only maintains privacy but also provides individual fish farmers with access to a broader dataset. By having access to more comprehensive data, than governance limitations dictates, fish farmers can make better-informed decisions, optimize their operations, and improve fish health and welfare. This method would also open AquaCloud insights to benchmarking and product fit evaluation for external actors.  

In this pre project, we prove that the method is applicable to AquaCloud and demonstrate how federated learning can improve model performance in the aquaculture industry based on AquaCloud data. The report describes the foundation for how a non-invasive platform can be implemented to yield flexible and secure collaborative model training. Implementing such infrastructure for federated learning have the potential to increase industry collaboration and generate significant value, driving advancements in fish health, operational efficiency, and overall sustainability.

Symmary of results

Blockquote: "Increased volumes of data generally leads to better models".

•Federated models outperform individual models globally

•Data sharing generally enhances local model performance

•Federated and conventional models have comparable performance

•Ensemble models have comparable performance for gradient boosting