Adversarial resilience in EEG-based BCI systems: a two-tiered approach using GANs and transfer learning

[EN] Brain-Computer Interface (BCI) technology holds immense potential for enhancing human life by decoding brain signals. However, its susceptibility to adversarial attacks remains a significant barrier to real-world adoption. Although recent works have employed Generative Adversarial Networks (GAN...

Descripción completa

Detalles Bibliográficos
Autores: Sayah Ben Aissa, Nour El Houda, Kerrache, Chaker Abdelaziz, Korichi, Ahmed, Lakas, Abderrahmane, Hernández-Orallo, Enrique|||0000-0002-3284-561X, Tavares De Araujo Cesariny Calafate, Carlos Miguel|||0000-0001-5729-3041
Tipo de recurso: artículo
Fecha de publicación:2025
País:España
Institución:Universitat Politècnica de València (UPV)
Repositorio:RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia
Idioma:inglés
OAI Identifier:oai:riunet.upv.es:10251/230336
Acceso en línea:https://riunet.upv.es/handle/10251/230336
Access Level:acceso embargado
Palabra clave:Electroencephalography (EEG)
Classification
Adversarial attacks
Generative adversarial networks
Transfer learning
Descripción
Sumario:[EN] Brain-Computer Interface (BCI) technology holds immense potential for enhancing human life by decoding brain signals. However, its susceptibility to adversarial attacks remains a significant barrier to real-world adoption. Although recent works have employed Generative Adversarial Networks (GANs) to synthesize adversarial EEG signals for augmenting training data, they rely on disjointed CNN-based architectures for adversarial detection and EEG classification. To date, to our knowledge, no unified architecture has been proposed that can simultaneously detect adversarial examples and classify EEG signals. This paper addresses this gap by introducing a novel two-level architecture. The first level leverages GANs, where the generator synthesizes adversarial EEG signals and the discriminator functions as an Adversarial Detection System (ADS) to identify adversarial patterns. The second level focuses on classifying normal EEG signals, using transfer learning to enhance efficiency by leveraging knowledge from the first level. Evaluated on a widely-used EEG dataset, our approach demonstrates superior performance in both adversarial detection and EEG classification compared to state-of-the-art methods.