How It Works

Keystroke & Mouse Dynamics

Building a profile

During enrolment, each attempt produces a feature vector. After all required samples are collected, the profile stores the mean and standard deviation of each feature across attempts:

mean[i]  = average of feature i across all enrolment samples
std[i]   = standard deviation of feature i across all enrolment samples

Matching: normalised Manhattan distance

For a new sample, the distance to a stored profile is:

distance = mean over all features of  |sample[i] − mean[i]| / max(std[i], floor)

The floor prevents division by near-zero standard deviations when a feature is very consistent:

Modality	Floor values
Keystroke dwell	15 ms
Keystroke flight	25 ms
Mouse movement time	30 ms
Mouse click dwell	15 ms
Mouse curvature	0.03

Why this works: features with low variance (consistent behaviour) have a small std, so even a small deviation produces a large normalised distance. Consistent features are effectively weighted more heavily.

Converting distances to confidence

Distances are converted to confidence percentages using scaled softmax:

confidence[i] = exp(−scale × distance[i]) / sum(exp(−scale × distance[j]) for all j)

The negative sign means smaller distances → higher confidence. The softmax ensures all confidences sum to 100%.

The scale factor (default: 2.0) is configurable per modality via the Admin Panel. A higher value sharpens the separation — the best match receives a larger share of confidence relative to second-best. A lower value produces a flatter distribution when profiles are close together.

Face Recognition

Feature extraction

A TinyFace model detects 68 facial landmarks. 16 geometric ratios are computed from these landmarks (eye widths, nose proportions, mouth width, inter-feature distances), all normalised by face size to remove scale dependence.

Matching: cosine similarity

The similarity between an enrolment vector e and a sample vector s is:

similarity = (e · s) / (|e| × |s|)

A similarity of 1.0 means the vectors point in exactly the same direction (identical features). A similarity of 0.0 means they are orthogonal (completely different).

Similarities are normalised so they sum to 100% confidence across all enrolled profiles.

Voice Biometrics

Feature extraction

The browser records audio and computes a short-time Fourier transform (STFT) to produce a spectrogram. From the spectrogram, Mel-Frequency Cepstral Coefficients (MFCCs) are extracted per frame — 13 coefficients capturing the spectral envelope of the voice.

The enrolment profile is the mean MFCC vector across all frames of the recording.

Matching: cosine similarity

Identification uses the same cosine similarity formula as face recognition, applied to the 13-element MFCC vectors.

Signature Dynamics

Feature extraction

Six scalar features are extracted from the stroke data:

Feature	How computed
Duration	`end_time − start_time` in seconds
Path length	Sum of Euclidean distances between consecutive points, normalised by canvas diagonal
Average velocity	`path_length / duration`
Peak velocity	Maximum instantaneous speed between consecutive points
Stroke count	Number of pen-down events (mouse/touch down without up)
Direction-change rate	Count of direction reversals per second of drawing time

Matching: weighted Euclidean distance

Features are normalised per dimension, then combined:

distance = sqrt( sum over all features of  w[i] × ((sample[i] − profile[i]) / scale[i])² )

where scale[i] is a per-feature normalisation constant tuned to the expected range of each measurement, and w[i] is a feature weight (all currently equal). Distances are converted to confidence using softmax.