Heuristic Optimization of RNN and LSTM Models for Detecting DeepFake Voice

Deepfake voice synthesis, which enables the artificial replication of human speech through deep learning and natural language processing, poses increasing risks to information security and digital trust. Detecting such synthetic voices remains a challenging task due to the high realism and variability of generated speech. This study proposes an enhanced Deepfake voice detection framework that integrates heuristic optimization algorithms with deep learning architectures to improve model performance. Specifically, Recurrent Neural Network (RNN) and Long Short-Term Memory (LSTM) models were optimized using nine heuristic and meta-heuristic algorithms, including Genetic Algorithm (GA), Particle Swarm Optimization (PSO), Differential Evolution (DE), and Jaya. The ADD 2022 dataset—comprising 11,536 samples (8,275 fake and 3,260 real)—was preprocessed into spectrogram-based numerical features for training and testing. Each heuristic algorithm performed hyperparameter optimization over learning rate, dropout rate, input size, and batch size to maximize detection accuracy. Experimental results demonstrate substantial performance gains, with RNN–PSO and LSTM–DE combinations achieving 99.0% and 97.2% accuracy, respectively—an improvement of over 20% compared to the baseline model. These findings indicate that heuristic optimization significantly enhances generalization and convergence efficiency in Deepfake voice detection. The proposed approach contributes a scalable optimization framework adaptable to other deep learning-based media authentication tasks.