In real-world anomaly detection, such as industrial and financial fraud detection, most data is unlabeled, with only a small portion of labeled anomalies. Semi-supervised anomaly detection methods based on reconstruction models are commonly used, but their performance is often limited by the scarcity of labeled data, which hinders learning of underlying patterns. To address this, we propose the Kernel Density Auto-Encoder (KDAE), which employs Kernel Density Estimation (KDE) to amplify significant features in each dimension, reducing the impact of limited data. During testing, KDAE leverages labeled data as supplementary knowledge to narrow the threshold range. The KDAE model achieves state-of-the-art performance on four real-world datasets, as demonstrated by ablation studies and sensitivity analyses.