Comparison of generative adversarial networks and diffusion models for synthetic steady-state visual evoked potential signal generation

Abstract

Steady-state visual evoked potential (SSVEP) signals are widely used in brain–computer interface (BCI) applications. Artificial generation of SSVEP signals can reduce the time required for experimental data collection, thereby minimizing user fatigue and improving comfort during electroencephalographic recordings. Synthetic SSVEP data enable data augmentation, training and validation of BCI classification algorithms, and benchmarking of SSVEP-based BCI systems without extensive subject-specific recordings. However, realistic modelling and generation of SSVEP signals remain challenging due to their nonlinear and non-stationary characteristics and the presence of harmonic components. In this study, we present a comparative analysis of two approaches to synthetic SSVEP signal generation: generative adversarial networks and diffusion models. The performance of both methods is evaluated using qualitative and quantitative metrics to assess the fidelity of the generated signals in both the time and frequency domains. The evaluation framework includes the frequency-domain Euclidean distance, Wasserstein distance, signal-to-noise ratio (SNR), and principal component analysis (PCA). The results indicate that diffusion models produce synthetic SSVEP signals with higher fidelity in both the time and frequency domains than the evaluated GAN configuration. Overall, diffusion models achieved better agreement with real data according to the adopted metrics while preserving key signal characteristics.