IEEE Consumer Electronics Magazine - January/February 2020 - 79

Figure 3. Sliding sampling and bagging.

In the feature selection process, this paper
does not select the id feature, as the id in the
data only indicates the trajectory number and
has no actual meaning. Finally, after feature
selection and data processing, the training data
are transformed into a 903-dimensional matrix
(including targets and labels).

Training and Bagging
The basic classifier selected in this paper is a
representative algorithm of ensemble learning,
i.e., xgboost. For real-world man-machine verification, this paper proposes a method of sliding
sampling and sliding bagging, which is shown in
Figure 3.
Sliding sampling: To solve the problem of data
imbalance, the model proposed in this paper
adopts the method of sliding sampling and merging datasets. The sliding sampling is after the
data processing. The main procedure is to separate the positive and negative samples then the
sample of large proportions is sampled according to the number of samples equal to a small
proportion. The sampled data are combined
with another type sample (in a small proportion)
and then transformed into a training set.
Sliding Bagging: Due to the multiple training
sets produced via sliding sampling, a large number
of models are trained on each sliding sample set,
bagging is then performed on the multiple models.
After performing data preprocessing, the classifier needs to be trained. According to the sliding
sampling procedure proposed above, the data
are divided into several training set, and each set
is used to train a base classifier. Then, the bagging approach is used, and all the trained classifiers are combined.
Conventional bagging in ensemble learning
involves a random sample, training multiple

January/February 2020

classifiers and ensemble, the sample of it is random with replacement. However, the sliding
sampling and bagging conducted in this paper is
different. The sliding sampling is likely to regular
without replacement. This article proposes a
new way that involves sliding sampling and combines multiple classifiers. The combination
method involves using each classifier to test all
the test data and weighting all the classification
results.

TESTBED FOR THE CASE STUDY
Testbed Architecture
In this article, the experimental machine consists of an Intel core i5 CPU at 1.6 GHz workstation running Windows 7 with 8 GB of RAM.
Finally, the entire CAPTCHA model is established
and appliedBaecher2014CAPTCHAs to the system, as shown in Figure 4.
The system structure is shown in Figure 4.
The whole system is a background program and
executing in the background process, and the
data acquisition is carried out by a background
process. At each time interval, the location of
the cursor and the relevant data are recorded.
Then, the data are input into the proposed
CAPTCHA model, which is based on cursor trajectories. After data preprocessing, data alignment,
feature selection, and processing are carried out,
and then sliding sampling is performed to train
the classifier. The test data are then classified.
Finally, the results are obtained via sliding
bagging.
After completing the whole process of the
model described above, it can be determined
whether a trajectory is a result of human or
machine operation. Using the collected data,
this paper conducts a relevant experiment.

IEEE Consumer Electronics Magazine - January/February 2020

Table of Contents for the Digital Edition of IEEE Consumer Electronics Magazine - January/February 2020