The rise of online social networks has heightened concerns over image privacy leakage. Although deep learning methods have been applied to privacy recognition, they face two key challenges: (1) a privacy gap between low-level visual features and high-level, context-aware human judgments, and (2) limited consideration of inter-entity context. To address these, we propose MSPP-Net, a Multi-Stage Privacy Perception Network inspired by human cognition. It decomposes privacy inference into three stages: entity perception to detect key objects, attribute perception to align visual features with semantic concepts via multimodal contrastive learning, and privacy perception to model inter-object context using graph attention networks. Experiments on our FineViP dataset show that MSPP-Net outperforms strong baselines, improving mAP by 3% and OR by 1.1%, validating the benefits of structured, cognitively motivated modeling for privacy recognition.