Statistical modeling of keystroke dynamics samples for the generation of synthetic datasets

Abstract

Biometrics is an emerging technology more and more present in our daily life. However, building biometric systems requires a large amount of data that may be difficult to collect. Collecting such sensitive data is also very time consuming and constrained, s.a. GDPR legislation in Europe. In the case of keystroke dynamics, most existing databases have less than 200 users. For these reasons, it is crucial for this biometric modality to be able to generate a significant and realistic synthetic dataset of keystroke dynamics samples. We propose in this paper an original approach for the generation of synthetic keystroke data given samples from known users as a first step towards the generation of synthetic datasets. Experimental results show the capability of the proposed statistical model to generate realistic samples from existing datasets in the literature.

Siamese Networks for Static Keystroke Dynamics Authentication

“…There is no pressure to collect several samples for enrollment as the network has been trained with other individuals samples and expect that few samples allow obtaining good performances with an increase of performance with the number of samples. To verify this point, we try different gallery size: [1,5,10,20,30,40,50,100,150,200]. We expect to obtain good performance even for few samples.…”

“…A generalization to a user-based password authentication [7] or even free text is expected. Artificial samples generation with handcrafted methods [9], [20] or Generative Adversarial Networks [21] could help to generate additional training data if required for such system.…”

My Behavior is my Privacy & Secure Password !

“…Fixed-text Keystroke Dynamics datasets used in this study are described in Table 1. As described in [5,6], these datasets have been cleaned and only the first 45 entries of each user are kept. Metrics given in this paper are computed as the average value of the metric across the 4 datasets.…”

“…However, such representation (raw) gives disappointing performances (EER=40%). As stated in [5], 6 duration times can be extracted from each digraph, with the last duration of a given digraph (d 5 ) also being the first duration (d 0 ) of the next digraph. However, as these duration times can be rewritten as additions of dwell and flight times, they are, by construction, not bringing any additional security or performance to the BioHashing algorithm.…”

“…A more practical estimation of cdf i (X i ) is to compute the parameters of the law X i is assumed to follow. In this study, we used the same laws (gumbel, normal, logistic, and laplace) and fitness functions (raw, R mle, R mge, and R qme) used in [5]. All dwell times were assumed to follow the same law, but with different parameters, as well for the flight times.…”