114 COMMUNICATIONS OF THE ACM | FEBRUARY 2017 | VOL. 60 | NO. 2
subject was varied between several hours and several weeks.
This was done in order to try to eliminate any effects of sampling at a specific time of the day.
Data extracted from the measurement setup is in the
form of a 4000 sample time-series describing voltage variation as seen by the oscilloscope. Figure 4 shows the input
pulse sent by the waveform generator and the pulse measured by the oscilloscope.
Time series measurements are converted to the frequency
domain using the FFT and the first 100 frequency bins of
the FFT data are used for classification. Operating in the frequency domain has several advantages. First, there is no need
to worry about alignment of the measured data pulses when
computing metrics, such as the Euclidean distance between
pulses. Second, it quickly became apparent that only lower
frequency bins carry any distinguishing power. Higher frequency bins were mainly noise, meaning that the FFT can
be used to perform dimensionality reduction of the original
4000 sample time-series to the vector of 100 FFT bins.
7. 4. Results
We present two different classifiers: one for authentication
and one for identification. The former is based on SVM and
verifies a 1 : 1 match between a sample from an unknown person and that of a requested person. The identification classifier, also based on SVM, verifies a 1 : n match between a sample
of a known person against all samples in a database. The identification classifier is of a closed-set variety. Section 2 provides
a more detailed description of open- and closed-set classifiers.
We sub-divide results into: ( 1) those from a single test-set,
which show the distinguishing power of pulse-response, and
( 2) those based on data sampled over time, which assess stability (permanence) of pulse-response.
Authentication classifier. Figure 5 shows the distinguish-
ing potential of the authentication classifier applied to a data
set collected over several weeks. Each bar shows the classi-
fier’s performance for different threshold levels, for each of
the test subjects. The threshold is a measure of assurance
of correct identification. If a low false positive rate is accept-
able, better sensitivity can be achieved. The classifier’s perfor-
mance is measured using fivefold cross-validation to ensure
statistical robustness. The figure shows that all subjects are
recognized with a very high probability, as the true positive
Applying the authentication classifier to the single-session
data set yields even better performance figures (see the full
version of this paper in Rasmussen et al. 8). For example, 10%
false positives allow us to achieve sensitivity of almost 100%.
Identification classifier. Identification is a multi-class
classification problem. Our classifier consists of multiple
SVMs and follows a one-against-one approach (aggregation
by voting). Due to this increased complexity a slight drop in
performance is expected, in comparison to authentication,
which is a binary classification task.
Results obtained from the identification classifier over
the two data sets are shown in Figure 6. Even with increased
complexity, the identification classifier performs very well
on both data sets. The single-session data set contains ten
people and the goal of the classifier is to identify each person as accurately as possible. There is a slight decrease in
performance for the data set containing samples taken several weeks apart. The reason for this decrease is that samples
taken far apart are influenced by very different conditions.
0 200 400 600 800
Figure 4. Input and output waveforms. One measurement consists
of 4000 samples with the rate of 500 MSa/s.
Figure 5. True positive rate for each test subject for the authentication
classifier fed with the data sampled over time. Error bars show 95%
confidence interval. The x-axis reflects the discrimination threshold
for assigning the classifier’s prediction output to a positive or a
90 92 94 96 98 100
Over time Single data set
Aiden Ethan Jacob LiamMason Charles David EthanJackson Liam LucasMason NoahRichardSophia
Figure 6. Identification classifier results. The true positive rate for
each test subject is obtained by applying five times stratified fivefold
cross-validation. Error bars show 95% confidence interval.