Interpretable Deepfake Voice Detection: A Hybrid Deep-Learning Model and Explanation Evaluation

With the unprecedented advancement of Generative Artificial Intelligence (GenAI), the threat of voice scams using synthetic voices has become a serious concern across various sectors. Recent efforts have focused on identifying fake voices through handcrafted features, deep learning models, and hybrid approaches. However, most existing methods lack explainability, rendering their predictions non-transparent to users. This paper proposes a novel, interpretable, and transparent method for fake voice identification by introducing a hybrid deep learning model that leverages multiple extracted features. The hybrid model consists of two main components: the first component addresses heterogeneous feature spaces by employing deep convolutional sub-models tailored to individual features, while the second component, the terminus model, utilizes the concatenated representations from the final layers of each sub-model as input. The terminus model follows a typical multi-layer perceptron architecture, enabling effective integration and classification of the diverse feature representations. To enhance interpretability, we decompose the model’s decisions using Local Interpretable Model-agnostic Explanations (LIME), taking advantage of the identical feature representation before the concatenation layers to address challenges related to multidimensional feature representations. To evaluate the features and assess the quality of the generated explanations, we propose two metrics: importance and trust. Extensive experiments are conducted on the In-the-Wild dataset, which is designed to test the generalization capability of synthetic audio detection methods. The experimental results demonstrate that our approach achieves performance comparable to benchmark methods. Furthermore, the results based on our proposed metrics conclude that certain perceptible features demonstrate promise for generating explanations that are meaningful to general users. For reproducibility, the source code for these experiments is available in the following repository: https://github.com/jacoblarock/fake_voices_xai