ARES: Adaptive robust object detection framework for enhancing real-time performance in autonomous vehicle systems

Abstract

In contemporary autonomous vehicles, object detection must provide both robust detection against threats like adversarial patch attacks and timely execution to meet real-time deadlines. Certifiably robust detection, also known as patch-agnostic approach, meets the first requirement. However, it introduces significant computational overhead, thereby compromising its real-time performance. To resolve this conflict, we propose ARES, a novel framework inspired by mixed-criticality systems. ARES introduces a security-driven paradigm. By default, the framework operates in a high-performance, low-security mode. However, it transitions to a high-security mode, utilizing a computationally intensive and robust detector, only when an active attack is detected. This selective activation is managed by the ARES transition manager, which captures the attack timing and handles tasks during mode transition. The ARES scheduling framework, on the other hand, guarantees formal schedulability analysis and optimal priority assignment. In our experiments, ARES demonstrated an increase of up to 8.9$\times$ in overall FPS detection over baseline. Furthermore, when evaluating the acceptance ratio with randomly generated task sets, ARES exhibited a 40.8–62.9% enhancement in schedulability compared to baseline.