Detecting Concept Drift in Object Detection Models: A Collaborative AI-Human Approach for Defense Applications

Abstract

Concept drift, a persistent challenge in machine learning, can lead to the deterioration of model performance over time due to changes in data distribution. This issue is a pressing issue for the U.S. Armed Forces as they integrate AI-driven systems into defense operations such as surveillance and reconnaissance as drops in model accuracy can lead to sever tactical disadvantages. This study presents a novel drift detection method that quantifies concept drift by combining AI-generated confidence values with human corrections. Unlike traditional approaches, this method allows for AI-human interaction, ensuring adaptability to evolving data distributions while enhancing trust and facilitating targeted retraining. The proposed system introduces a custom metric called the Drift Avoidance Value to detect drift in desired object classes. A YOLO object detection model was employed to evaluate the method using a dataset of 200 images focusing on object classes relevant to military operations. The experimental results confirm the effectiveness of the Drift Avoidance Value for detecting and responding to concept drift, decreasing as the model's performance declines and increases following retraining. The study highlights the crucial role of human input in reinforcing model reliability and demonstrates the method's effectiveness in adapting to performance degradation. This approach supports the sustainability of object detection systems in defense and could also be applied to various other AI systems, ensuring robust decision-making across multiple domains.