Dynamic Confidence-Aware Fusion for Multi-Modal Anomaly Detection in Autonomous Systems

In the current scenario, autonomous platforms like self-driving cars and drones are relying more and more on sensor-driven decisions. But making sure they are secure and can detect problems in time is still a tough task. Traditional machine learning models, especially ensemble ones, often don’t hold up when data is messy, constantly changing, or under attack.This paper introduces a practical framework called Dynamic Confidence-Aware Fusion (DCF). It’s based on Bayesian theory and adjusts how much each model’s opinion is trusted, depending on how confident it is at the time. We’ve tested DCF using a simulation built in Python, where we fused data from LiDAR, radar, vision, and V2X sources. Our results showed strong F1-scores: 97.8% on clean inputs, 91.4% under adversarial attacks, and 88.2% when the data was out-of-distribution. The latency was low too—just 45ms. All of this suggests that DCF is a solid step forward in real-time, trustworthy anomaly detection.