Scaling Synthetic Training Data for Mouse-Dynamics Authentication

Per-user authentication from mouse dynamics degrades when a user has contributed little enrolment data, motivating the use of generative models to synthesize additional genuine sequences. We evaluate whether TimeGAN-synthesized sequences improve a per-user BiLSTM authenticator across 14 users, using a pipeline in which the generator and the authenticator consume identical preprocessed sequences. Synthetic sequences reduce Equal Error Rate (EER), and the reduction scales with how many are added: from 0.0093 at a 1:1 real-to-synthetic ratio to 0.0149 at 1:5, where 13 of 14 users improve (p = 0.002). Mean EER falls from 0.3416 to 0.3267. Returns diminish beyond a 1:3 ratio. Projecting sequences into the trained embedder's latent space shows that additional synthetic data increases density within regions the generator already covers, while regions of the real distribution left unoccupied at 1:1 remain unoccupied at 1:5. What limits further improvement therefore appears to be the generator's coverage, not the volume of data drawn from it