Global Navigation Satellite Systems (GNSS) are vulnerable to both suppressive and spoofing interference in complex electromagnetic environments, while traditional identification methods exhibit limited performance in dynamic scenarios. To address this, a novel identification framework integrating deep feature enhancement with a lightweight classifier was proposed. First, mathematical models and a mixed sample set for seven typical interference types were systematically constructed. Then, a feature enhancement module based on a long short-term memory network with an Attention mechanism (LSTM-Attention) was proposed. This module adaptively learned dependencies from raw features, focused on key information, and generated more discriminative 20-dimensional deep features. Finally, a structurally simple multi-layer perceptron (MLP) was employed for efficient classification. Experimental results show that the proposed method achieves an accuracy of 96.1% and an F₁-score of 96.0% on an independent test set, significantly outperforming various traditional machine learning and complex deep learning models. Its effectiveness and interpretability are further confirmed through confusion matrix analysis and feature importance evaluation. It demonstrates that, through deep feature enhancement, a classifier of moderate complexity can enhance computational efficiency while maintaining high accuracy, offering a feasible approach for the practical application of GNSS anti-jamming techniques.