Abstract
Machine Learning in coalition settings requires combining insights available from data assets and knowledge repositories distributed across multiple coalition partners. In tactical environments, this requires sharing the assets, knowledge and models in a bandwidth-constrained environment, while staying in conformance with the privacy, security and other applicable policies for each coalition member. Federated Machine Learning provides an approach for such sharing. In its simplest version, federated machine learning could exchange training data available among the different coalition members, with each partner deciding which part of the training data from other partners to accept based on the quality and value of the offered data. In a more sophisticated version, coalition partners may exchange models learnt locally, which need to be transformed, accepted in entirety or in part based on the quality and value offered by each model, and fused together into an integrated model. In this paper, we examine the challenges present in creating federated learning solutions in coalition settings, and present the different flavors of federated learning that we have created as part of our research in the DAIS ITA. The challenges addressed include dealing with varying quality of data and models, determining the value offered by the data/model of each coalition partner, addressing the heterogeneity in data representation, labeling and AI model architecture selected by different coalition members, and handling the varying levels of trust present among members of the coalition. We also identify some open problems that remain to be addressed to create a viable solution for federated learning in coalition environments.
Introduction
In modern military coalitions, such as those involving the U.S. and the UK, the ability to leverage artificial intelligence at the "tactical edge" is increasingly critical. Sensor data analysis and threat prediction require robust machine learning models. However, traditional centralized learning approaches—where data is gathered in a single location for training—are often untenable in coalition environments.
Three primary obstacles hinder centralized learning: security and privacy concerns prevent partners from sharing raw sensitive data; resource constraints like low bandwidth and limited power make transferring massive datasets impossible; and heterogeneity in equipment and terminology leads to significant "Linguistic and Lexicon Mismatches." This research, conducted under the Distributed Analytics and Information Science International Technology Alliance (DAIS ITA), proposes a Federated Learning (FL) framework designed to overcome these barriers by moving the learning process to the data.
Federated Learning Architectures
The proposed framework categorizes coalition learning into two primary modes based on operational constraints and trust levels:
- Data Sharing Mode: Partners share summarized versions of their data, such as "coresets" or "sketches," to a central server. This mode is suitable when bandwidth is sufficient and privacy policies allow for the exchange of data summaries.
- Model Sharing Mode: This more sophisticated approach involves partners training models locally on their private datasets and exchanging only the resulting model parameters (weights). This minimizes bandwidth usage and maximizes data privacy.
Policy-Driven Interoperability
A significant challenge in coalition operations is the lack of uniformity in data representation. One partner might label a vehicle as a "Truck," while another uses the term "HGV." To address this, the paper introduces a Generative Policy Framework.
Instead of relying on static, hard-coded rules, the system utilizes "Answer Set Grammars" to automatically generate operational policies. These policies manage data translation (e.g., format conversion), relabeling features to a common schema, and establishing thresholds for the Quality of Information (QoI). This automation allows for agile coalition setup and ensures that data from diverse sources can be effectively integrated without manual intervention.
Addressing Data Skew and Model Fusion
In tactical environments, local datasets are often highly skewed. For instance, one site might only observe tanks while another sees only civilian vehicles. In such cases, simple weight averaging—a common technique in standard Federated Learning—fails to produce an accurate global model.
The authors propose an ensemble-based fusion approach using a "policy-based selector." This method recognizes that models have "applicability boundaries." Rather than forcing a single master model, the system determines which local model is most relevant to a specific region of the feature space during inference.
This approach ensures that a model trained on low-resolution desert footage is not inappropriately applied to high-resolution forest analysis, thereby maintaining the integrity and precision of the overall system.
Conclusion and Future Directions
While the proposed framework addresses many of the core challenges of coalition AI, several open problems remain. Current FL methods often require synchronous communication, which is difficult to maintain in active warzones where units frequently go offline. Asynchronous Federated Learning remains a critical area for future research. Additionally, while this work focuses on the Learn phase of the AI lifecycle, extending these principles to Federated Inference and Federated Action is essential for achieving fully coordinated coalition responses. By bridging the gap between local data privacy and global model intelligence, this framework provides a foundation for more effective and secure coalition operations.
Summary updated