To test the model beyond the testing set provided by us, you will be creating
new fake signals (attack vectors). There is no need to do this on the phone,
and you can do it on a machine. There are three general approaches, but you
are not limited to them. You need to create at least two new attack vectors,
one should be using signal generation methods (check 6.b), and the other is
of your choice.
a. Noise Addition: You will add noise to original signals. Noise can be
added in time domain, or in feature domain (e.g. frequency, wavelet,
..). If you decided to add noise in feature domain, you need to make
sure the feature extraction method is reversible, so you can get back a
new time domain signal.
b. Signal Generation: One can train generative/predictive models to
generate new signal. Any model used for time-series
prediction/forecasting can also be used. Some models are ANFIS,
GAN, VAE, ....
c. Random Inputs: Create random signals in time domain, or random
vectors in feature domain, and then map them back to time domain.

a. You will be having at least five trained machine learning models (check
4) on your system and will report the decision (live or fake) of each
model on the phone, and afterwards do a voting on models, and report
the final decision, and the true label of the input.
b. The app will be having two modes; A) one single input is received and
the operation is as described in 3.a, and B) a group of input is received,
which then you will report the aggregate results (accuracy, false accept
rate, false reject rate, half total error, F1 score, ...) for each of the
models, and for the voting model.

Create an app capable of receiving data through Internet or Bluetooth. I would
suggest using Internet option since then you can easily test your system using
any machine, and not constrained by having Bluetooth capability. For
establishing internet communication, you can use a client server approach
which standard libraries are available for it.