Centrally Pretrained Federated Fine-Tuning: Enabling a Secure and Accurate Military Security Application on Embedded Hardware
Abstract
A persistent, precise, and adaptive security application is a requisite component to an effective force protection condition (FPCON) as U.S. military installations have become common targets for violent acts of terrorism and homicide. Current military security applications require a more automated approach as they rely heavily on limited manpower and limited resources. The current research developed an off-grid, deployed federated fine-tuning network composed of embedded hardware and evaluated embedded hardware system and model performance. Federated fine-tuning takes a centrally pretrained model and performs fine-tuning on a select number of model layers within a federated learning architecture. The federated fine-tuning models exhibited an average reduction in CPU load of 65.95 percent and an average reduction in current draw of 56.18 percent. The MobileNetV2 model transmitted 81.59 percent fewer global model parameters across the network. The centrally pretrained MNIST model began training with an initial accuracy improvement of 53.94 percent over the randomly initialized model. The centrally pretrained MobileNetV2 model demonstrated an initial average accuracy of 90.75 percent at training round 0 and experienced a 3.14 percent overall performance improvement after 75 federated training rounds. The results of the current research demonstrated that federated fine-tuning can improve system performance and model accuracy while providing stronger privacy and security against federated learning attacks.
Document Details
- Document Type
- Technical Report
- Publication Date
- Dec 01, 2020
- Accession Number
- AD1126761
Entities
People
- Matthew W. Baxter
Organizations
- Naval Postgraduate School