With the popularity of online shopping in recent years, various prohibited items are continuously attacking e-commerce portals. Searching and deleting such risk items online has played a fundamental role in protecting the health of e-commerce trades. To mitigate negative impact of limited supervision and adversarial behaviors of malicious sellers, current state-of-the-art work mainly introduces heterogeneous graph neural network with further improvements such as graph structure learning, pairwise training mechanism, etc. However, performance of these models is highly limited since domain knowledge is indispensable for identifying prohibited items but ignored by these methods. In this paper, we propose a novel Knowledge Based Prohibited item Detection system (named KBPD) to break through this limitation. To make full use of rich risk knowledge, the proposed method introduces the Risk-Domain Knowledge Graph (named RDKG), which is encoded by a path-based graph neural network method. Furthermore, to utilize information from both the RDKG and the Heterogeneous Risk Graph (named HRG), an interactive fusion framework is proposed and further improves the detection performance. We collect real-world datasets from the largest Chinese second-hand commodity trading platform, Xianyu. Both offline and online experimental results consistently demonstrate that KBPD outperforms the state-of-the-art baselines. The improvement over the second-best method is up to 22.67% in the AP metric.