Increasing demand for high performance and energy efficiency along with time-to-market pressures have led to a growing number of third-party accelerators in modern System-on-Chips (SoCs). However, the third-party accelerators may introduce vulnerabilities due to malicious intent or elusive design bugs that escape through verification processes. These vulnerabilities can be exploited by various attackers which can corrupt the entire system. Flooding is one such attack that is triggered by a malicious third-party accelerator to obstruct on-chip network communication by injecting useless packets leading to Denial-of-Service. To secure accelerator-rich SoCs from flooding attacks, we propose a two-step attack detection framework that leverages machine learning (ML) models to detect an attack scenario. The ML-based solution along with hierarchical feature data aggregation and feature packet prioritization provides an accurate attack detection with minimal performance overheads. Experimental evaluations with real accelerator benchmarks show the effectiveness of our framework achieving a detection accuracy of up to 97 percent across various ML classifiers.