HYPER-AUTH: a hybrid PCA-encoder recurrent architecture for EEG-based authentication

Creative Commons License

Cyrille T. N. R., Jimenez A. M. O., Balli T., Yetkin E. F.

Engineering Research Express, vol.8, no.18, 2026 (ESCI, Scopus)

  • Publication Type: Article / Article
  • Volume: 8 Issue: 18
  • Publication Date: 2026
  • Doi Number: 10.1088/2631-8695/aea48b
  • Journal Name: Engineering Research Express
  • Journal Indexes: Emerging Sources Citation Index (ESCI), Scopus
  • Keywords: biometric security, dimensionality reduction, EEG-based authentication, GRU autoencoder, PCA, SVM classification
  • Open Archive Collection: AVESIS Open Access Collection
  • Middle East Technical University Affiliated: Yes

Abstract

Biometric authentication relies on unique physiological traits such as fingerprints, facial features, iris patterns, or voice for secure identification. Among these unique characteristics, electroencephalography (EEG) signals offer distinctive neural signatures that make user authentication reliable and difficult to spoof compared to conventional biometrics. However, converting raw, non-stationary EEG waveforms into reliable features requires advanced processing that is highly susceptible to temporal data leakage. In this paper, we present a systematically evaluated two-stage pipeline: In the first phase, a GRU autoencoder operating in full 32-bit precision compresses overlapping 5 s EEG windows into 128-dimensional latent codes; then, principal component analysis (PCA) selects (Formula presented) (Formula presented) 84 principal components that explain 99% of the variance. The second phase makes use of a radial basis function support vector machine. Crucially, we implement a strict zero-leakage evaluation protocol using trial-aware group splits, ensuring that statistical scaling, PCA projection axes, and model hyperparameters are never exposed to test-trial data. Evaluated on 15,695 segments from 109 subjects (PhysioNet Motor Execution), the system yields a generalized accuracy of 85.93% and a macro (Formula presented) (Formula presented) -score of 0.86 on entirely unseen trials. Of the 109 users, 23 achieved an (Formula presented) (Formula presented) -score of at least 0.97, with 16 identifiable edge cases ( (Formula presented) (Formula presented) ) highlighting areas for targeted refinement. While overlapping-window leakage artificially inflates accuracies to near-perfect levels in naive implementations, our 85.93% true generalization rate demonstrates the real-world viability of these neural features for identity verification. Furthermore, the lightweight GRU pipeline has an inference time of approximately 145 ms per each 5 s segment window on standard CPU hardware, comfortably satisfying sub-second latency constraints and striking an effective balance between representational power and computational efficiency required for continuous authentication deployments.