Programming education continues to face significant challenges, with failure and dropout rates exceeding 30% in many introductory courses. Existing learning analytics approaches largely rely on static behavioral indicators derived from the Felder–Silverman Learning Style Model (FSLSM), which often fail to capture the temporal dynamics of learning and the syntactic complexity involved in programming activities. These limitations are particularly evident in detecting the Sequential/Global learning dimension and understanding how students interact with programming tasks over time. This study aims to address these limitations by proposing CAFNet (Crossmodal Attention Fusion Network), a multimodal learning analytics framework that integrates behavioral machine learning with natural language and code analysis. The proposed architecture combines Temporal Convolutional Networks to model behavioral indicators, CodeBERT for forum discourse representation, and Tree-Transformer models for Abstract Syntax Tree-based code analysis. A hierarchical cross-modal attention mechanism aligns these heterogeneous data sources, while Federated Supervised Contrastive Learning ensures privacy-preserving deployment across institutions under differential privacy constraints (ε = 0.5). The framework was evaluated using three heterogeneous datasets comprising 14,308 learners from programming education environments. Experimental results show that CAFNet achieved 91.7% classification accuracy with an AUC-ROC of 0.947, outperforming classical machine learning and deep learning baselines by 17.5%. The model achieved 94.1% accuracy for the Sequential/Global dimension, representing a major improvement over previous studies. Additionally, early at-risk prediction reached 88.9% accuracy at week four of the course. These findings demonstrate that integrating behavioral, linguistic, and programming data provides a scalable and privacy-compliant approach for intelligent educational systems supporting personalized learning and early academic intervention.