Federated Inference Using Self-Generated Policy
Military / Coalition Issue
When operation at the tactical edge or in any distributed environment, it may not always be possible to share data across sites or edge nodes. One approach to address this limitation is to federate the results of the model inference across site. In this demo, we will show a rule-based approach of federated inferencing, where the rules are self-generated based on contextual conditions.
Core idea and key achievements
- Inferencing across the tactical edge when data cannot be shared for variety of reasons, e.g., Security/Reliability and Network Bandwidth.
- Peer to peer connectivity to edge nodes when connectivity to central site of cloud is limited.
- Dynamic policy generation of classification ensemble weights to alleviate the need for end user to manually author and edit policies.
- Classification of events or data using ensemble approach across distributed edge nodes.
Implications for Defence
The key achievements in the previous section have implications in military defense settings as well as civilian. Classification of data, without connectivity to a shared or cloud server, may be challenging in an edge setting such as a military base. In some cases, trained models may not be mature enough to accurately classify certain data. Federated Inferencing can help in some cases by reaching out to nearby coalition bases to find more mature models to improve the accuracy of machine based classification.
Readiness & alternative Defence uses
Version 1 of the Edge AI SDK that incorporates all the described functionality is currently available within IBM Research. Ready to prototype a military Edge AI production facility, and experiment with a range of potential usage. Deployed in an experimental 5G Edge testbed at IBM Research.
Resources and references
- Verma, Dinesh, Seraphin Calo, and Greg Cirincione. “Distributed AI and security issues in federated environments.” Proceedings of the workshop program of the 19th International conference on distributed computing and networking. 2018.