Two-level keystroke system achieves 96% fixed-text accuracy and 83.5% TAR for continuous auth

The five-minute version

The two-level system achieves 3.77% EER on fixed-text login and 10.05% EER (83.53% TAR @ 5% FAR) on free-text continuous authentication, demonstrating low error rates. Level 1 uses an Evo-BioTCN with population-based training to adapt to long-term behavioral drift, while Level 2 monitors typing with a four-model deep-learning ensemble and re-authenticates on drift detection. Evidence comes from the GREYC (133 users, 110 days) and Buffalo (75 users) datasets, compared against static baselines such as Scaled Manhattan Distance and Transformer models. However, the reported 5% false-acceptance rate and evaluation only on zero-effort impostors temper the claim of strong resistance. The paper notes limitations in device diversity, user count, lack of targeted attacks, and computational cost of the ensemble.

What they found 2

• Fixed-text authentication achieves 3.77% EER and 96.87% TAR @ 5% FAR.

  “The system achieved an EER of 3.77% and a TAR of 96.87% @ 5% FAR.”

• Free-text continuous authentication achieves 10.05% EER and 83.53% TAR @ 5% FAR.

  “The final system achieves 96%+ accuracy for fixed-text authentication (3.77% EER) and 83.53% True Acceptance Rate (TAR) at 5% False Acceptance Rate (FAR) for free-text continuous authentication (10.05% EER)”

What undercuts it the paper’s own results

• The reported false acceptance rate of 5% may be considered high for high-security scenarios.
• The evaluation assumes zero-effort impostors (random attacks), not targeted adversarial attempts.

Stated limits 4

• The evaluation relies on the GREYC [14] and Buffalo [15] datasets. While GREYC [14] is longitudinal, it contains only two keyboard types, limiting the assessment of cross-device hardware variability (e.g., switching from mechanical to membrane keyboards).
• The free-text evaluation is limited to 75 users.
• The evaluation assumes zero-effort impostors (random attacks). Targeted adversarial attacks, where an impostor uses generative AI to mimic a specific user’s rhythm, were not empirically tested.
• The 4-Model Ensemble requires running four parallel deep networks, which may pose latency challenges on resource-constrained edge devices compared to single lightweight models.

What’s new here 4

• Introduces a two-level architecture combining fixed-text login with continuous free-text monitoring.
• Applies population-based training to evolve lightweight adapters for drift adaptation without catastrophic forgetting.
• Uses a four-model deep learning ensemble with diverse metric learning losses for free-text authentication.
• Integrates a Trust Anchor loss to mitigate slow-mimicry (frog-boiling) attacks.

Artifacts & methods 7

• Evo-BioTCNsystem
• Population-Based Training (PBT)
• Temporal Convolutional Network (TCN)primitive
• 4-Model Deep Learning Ensemblesystem
• Trust Anchor loss
• GREYC datasetdataset
• Buffalo datasetdataset

Topics