JoWUA

Abstract

The paper proposes an advanced federated hybrid deep learning framework combined with an attention mechanism for personality prediction based on handwriting documents. This framework utilizes CNN to extract detailed spatial features from handwriting, BiLSTM networks to capture dynamic temporal stroke features, and an attention mechanism that highlights psychologically relevant handwriting features. Differing from conventional central processing methods, which require the storage of sensitive handwriting data on the server, this method takes a step forward and uses federated learning. By learning a model over multiple distributed clients and transferring the model weights, the framework allows the privacy and security of the personal handwriting information without it being transferred to the server. The attention mechanism enhances the model accuracy and interpretability by adaptively allocating weight to each critical feature, including handwriting slant angle, spacing habits, baseline trend, and stroke continuity. According to experimental results, it achieves 94.2% accuracy, 92.8% precision, 93.1% recall, and 92.9% F1-score, which outperforms CNN (87.5%), RNN (88.3%), and centralized CNN-RNN (91.8%)-based models. Ablation study results indicated that BiLSTM captures effective sequential feature learning, the attention mechanism effectively learns discriminate features, and federated learning is capable of delivering high prediction performance along with privacy protection. In conclusion, the hybrid deep learning, attention mechanism, and federated learning combination is a feasible, explainable, and privacy-protected technique for handwriting-based personality assessment, providing a solid ground for secure psychological assessment.