Encrypted Model Inference for Secure ML Deployment

Artificial Intelligence (AI) and Machine Learning (ML) models are susceptible to adversarial attacks and, to address this threat, Project Linchpin is focused on integrating capabilities that safeguard model and data integrity. Ensuring model and data integrity in a Zero Trust Architecture means protecting AI/ML models, the insights extracted, and the associated training data, under the assumption that the network/platform has been breached. Enveil specifically engineered its ZeroReveal Machine Learning solutions to address AI/ML vulnerabilities and enable decision dominance within a multi-domain operating environment. EnveilÆs ZeroReveal Machine Learning Encrypted Evaluation (ML-EE) solution encrypts models during processing, protecting the model itself as well as the insights derived. By leveraging homomorphic encryption and secure multiparty computation, key pillars of the Privacy Enhancing Technologies category, ZeroReveal ML-EE ensures critical information ¢ such as model parameters and mission-sensitive data used during training ¢ is never exposed, which ultimately protects the interests and intent of Army mission owners. These capabilities would enable the Army to leverage AI/ML models to securely and privately derive insights from data sources across security and organizational boundaries, even when using highly sensitive or proprietary models, including those trained on classified or mission-sensitive data.