Camera source identification is widely used in digital forensics, particularly in copyright enforcement and tampering analysis. However, the compression and post-processing of images during social network transmission often lead to quality degradation, which adversely affects the accuracy of camera source identification. To tackle this problem, a Multi-scale Feature Fusion Integrated Attention Network (MFIAN) is proposed. First, an Integrated Attention Mechanism (IAM) is designed to combine the advantages of three classical attention mechanisms, and it is applied to multi-scale feature extraction, forming a W-shaped network (WNet-IAM). In addition, to effectively fuse low-level, mid-level, and high-level semantic features, a feature fusion network (MSLFNet) is developed. Experiments on three datasets show that the proposed approach boosts camera source identification accuracy by about 11% for images posted on social media, and enhances robustness against image processing with at least a 5.82% increase in accuracy. Ablation studies further verify that both IAM and MSLFNet significantly contribute to performance improvements, confirming the effectiveness of the proposed fusion network.