Against the backdrop of IoT-driven digital transformation in e-commerce, fake reviews have increasingly disrupted user decision-making and market order. Analyzing the credibility of user-generated content has become critical for maintaining a healthy digital ecosystem. To address this challenge, this paper proposes a multimodal fake review detection framework based on CB-ALBERT. First, the Cross-modal Contradiction Perception Attention Module (CPAM) explicitly captures contradictory signals between textual descriptions and metadata statistical features using a learnable L2 norm difference metric. Second, the Bidirectional Residual Dynamic Gating (BRDG) mechanism constructs bidirectional residual pathways for text and metadata, dynamically adjusting cross-modal feature fusion ratios through gating weights. Experiments show that the framework achieves an F1 score of 0.916 in e-commerce scenarios, representing a 4.5% improvement over mainstream models. While validated with e-commerce data, its multimodal contradiction awareness and dynamic fusion strategies can be directly applied to IoT scenarios, providing universal technical support for cross-domain fake information governance.