Advanced Persistent Threats(APTs) are targeted, stealthy, and highly sophisticated cyberattacks posing serious risks to critical infrastructure and sensitive data. In recent years, researchers increasingly focus on constructing provenance graphs from system logs to represent information flows, aiming to detect APTs through such semantically rich structured data. Although provenance graph-based APT detection techniques have shown effectiveness, they still suffer from several limitations: First, most of these methods require real-world APT data and prior expert knowledge; Second, they generally focusing on fine-grained local information while neglecting global information, resulting in in-sufficient extraction of rich contextual information and thus high false positive rates; Third, their detection performance degrades against evasion strategies, indicating insufficient robustness.To address these problems, this paper proposes Advanced Persistent Threat Detection via Graph Contrastive Masked Autoencoder (APT-GCM), a self-supervised learning-based detection model. APT-GCM leverages a graph contrastive masked autoencoder to learn node representations from benign provenance graphs, transforming malicious nodes detection into an outlier detection problem. It integrates the advantages of masked graph autoencoder and graph contrastive learning for extracting both local and global graph information through a dual-branch structure, learning deep implicit features of provenance graph nodes. Additionally, a uniformity loss is introduced to optimize the embedding distribution and improve discriminability. To further improve robustness, a complementary view augmentation module is incorporated, which constructs a complementary view of the masked view and performs feature alignment of masked nodes, strengthening the model’s ability to extract stable semantic information from diverse perspectives and thus improving its robustness against evasion attacks. We evaluate APT-GCM on three sub-datasets from the widely used DARPA E3 dataset, and experimental results demonstrate that APT-GCM outperforms state-of-the-art detection methods.