The rapid spread of fake news on social media and digital platforms poses severe threats to public
opinion, democratic processes, and societal trust. Traditional detection methods relying on handcrafted
features or standalone deep learning models often struggle with evolving misinformation tactics,
linguistic variations, and multimodal content. This paper proposes a hybrid two-stage evolutionary
intelligence framework for robust fake news detection. In Stage 1, a Genetic Programming-based symbolic
regressor evolves compact mathematical expressions for feature combination and initial classification.
In Stage 2, Differential Evolution (DE) optimizes hyperparameters of a hybrid LSTM-Transformer
classifier, enhancing generalization and reducing overfitting. Textual features are extracted using
BERT embeddings, augmented with propagation patterns and user metadata. The framework achieves
low-latency inference (<80 ms per article) on standard hardware. Evaluation on benchmark datasets
(LIAR, FakeNewsNet, ISOT) and a custom multimodal dataset demonstrates superior performance:
accuracy 97.2%, precision 96.8%, recall 97.4%, F1-score 97.1%, outperforming standalone DL models
and recent evolutionary hybrids. The system offers explainable decisions via evolved expressions and
robustness to noise, sarcasm, and short-text challenges, promoting reliable digital media verification.
