In distribution networks, a high impedance fault (HIF) exhibits weak characteristics, often indistinguishable from normal disturbances, making their detection challenging. Traditional indicator threshold methods, which are empirically calibrated, suffer from poor adaptability and lack sensitivity when confronted with complex environments. To address these limitations, a novel method for detecting HIF in distribution networks based on extreme gradient boosting (XGBoost) is proposed. This method avoids the complex threshold tuning. First, an equivalent model of a 10 kV medium-voltage distribution system with HIF is established to obtain zero-sequence current data for both HIF and normal disturbances. On the basis of data normalization, XGBoost is employed to directly learn the mapping relationship between the raw measurement information and the HIF from the original data, thereby constructing an HIF detection model to minimize errors caused by feature extraction. Finally, extensive simulation results demonstrate that the proposed detection method exhibits superior sensitivity and rapidity in identifying HIFs while demonstrating strong generalization capabilities in scenarios involving noise and data incompleteness.