New Method Boosts AI Perception Model Robustness Against Attacks

AI perception models often struggle when deployed in new environments or face adversarial attacks, leading to degraded accuracy. Current methods for combining these models, like majority voting, are vulnerable to coordinated errors and don't effectively recover performance under distributional shifts. This paper proposes a new neurosymbolic approach that learns to detect errors without requiring prior domain knowledge. The core innovation lies in exploiting vector-space geometry. By building Label Vector Pools (LVP) from each model's training embeddings, the system can derive error-detection rules based on the geometric relationship of detections to training prototypes. This method achieves performance comparable to domain-knowledge-based rules and significantly enhances robustness. The fusion process is framed as an abduction problem, solved using an Integer Program or a polynomial-time heuristic. Tested on aerial imagery with weather-shifted datasets and adversarial attacks, the approach maintains performance under label-flipping attacks, showing a 22% relative gain over majority voting at high flip rates.