Keystroke HMM Authentication

Behavioral authentication using keystroke timing from a fixed password.
We discretize timing features and train a discrete HMM per user, then identify the most likely user by log-likelihood.

Data

• Keystroke Dynamics Benchmark (CMU CyLab), DSL-StrongPasswordData.csv
• Each row = one repetition of the same password by a subject
• Features include hold times (H.) and latencies (DD., UD.*)

Approach

1. Preprocess: take the 31 timing features and discretize them into m bins (quantile-based).
2. Train: one discrete HMM per user with Baum–Welch (EM).
3. Predict: for a new sequence, compute log-likelihood under every user’s HMM and choose the max.

Notes

• Discrete HMM requires discretization of continuous timing values.
• Results depend on hyperparameters like number of bins m and number of hidden states.