In this paper, a Hardware Trojan (HT) detection model called BGNN-HT based on bidirectional graph neural network is proposed, which can detect HT cells by assessing the structure of its surrounding cells at gate level. BGNN-HT can precisely detect HT cells in circuits, and it does not require the golden model or manual feature extraction, which greatly reduces the difficulty of detection and can adapt to unknown HT. Experiments conducted on Trust-hub benchmarks show that when detecting unknown circuits and HTs, BGNN-HT can reach 96% True Positive Rate and 99% True Negative Rate in various datasets.