Lorica: AI agent microservices for federated data security

Eduworks Corporation (Eduworks) proposes to safeguard the Army’s AI data supply chain, both upstream and during model creation and deployment, with an application called Lorica.  Lorica will monitor the datastreams used by Army AI services, during training and at inference, and will observe the output of AI models. It will then alert analysts to suspected data poisoning, data manipulation, and unexpected changes in model behavior that could signal adversary actions and represent potential threats.  Lorica is a DoD-specific application of our innovative AI architecture and technology, called THIA. THIA uses a self-regulating microservice architecture to implement applications that (a) outperform the current state-of-the-art on reasoning and recall; (b) have low computational demands and can therefore run at the edge of a network on field equipment; and (c) can quickly adjust to new situations, automatically or with small amounts of input from subject matter experts. It includes optimized datastream interpreters for multiple types of data, including language, SQL, sensor, and multimedia data, and that can interpret the data in terms of high-level concepts.  Lorica will be an instance of THIA that has been trained to recognize potential manipulation and threats and that has a dashboard through which Army analysts will interact, receive alerts, and train Lorica without requiring prompt engineering or AI expertise. Lorica’s core model will gather input from microservices that look for drift, malicious data, and malicious intent in datastreams used by Army AI models during training and at runtime and that detect unusual behaviors in external models in Army AI systems. The Lorica core model will be a federation of small models, each of which employs a patent-pending technique that allows models to learn from experience without requiring re-training. The micromodels can be customized to different users or use cases without data being added to the context window. As Lorica operates, the data it retrieves from Army systems and the outputs of its internal microservices will be monitored to ensure they are in-domain, sufficient, and accurate in a way that is far more secure and less easy to attack than the guardrails used in large LLMs and in systems that use classical Retrieval Augmented Generation (RAG). The core model will examine the data received from Army systems to detect and alert potential manipulation or attacks.