ABSTRACT

oThis paper reviews the issue of deep fake audio

classification and evaluates the current methods of deep fake
audio detection for forensic linguist.
oAudios features were extracted
oPresented using MFCC, Mel-spectrum, Chromagram and
spectrogram
oFive iterative tests were used to determine the mean
accuracy
oResults showed a Custom Architecture
 INTRODUCTION

olacks of security
oPoses challenges for current biometric technology
oFew categories of risk of using biometrics
oNeed to get familiar with advances in new technology for
Digital forensic experts
oDrawbacks of new voices biometrics system
oInherent lack of techniques to identify voice deep fake
 CONTI…

oCybercrimes
oAnalysis process is the major component of digital forensic
 BACKGROUND
AND
RELATED WORKS

Deep Fake Audio and CNN architecture

Audio Deep fakes, scientifically know as logical access audio spoofing
methods, have become an enhanced challenge to voice interfaces due to
rapid breakthroughs in speech recognition and voice transfer technology.
Deep Fake deliver automated procedures to create false information tat is
more difficult for human analysts to spot.
 CONTI….

o Development of Deep fakes

the development of deep fakes are based on two machine learning progress
 Neural Networks
Represent how the human brain works, and they demonstrate how the brain
processes information. The faster and more precisely the brain can replicate the
more the human brain is introduced to representations of something, such as how
to catch a hit cricket ball or harmony for a new song.
 Generative Adversarial Networks
Deep fakes would not have been as credible without GANs as GANs consists of
two networks:
 The generator
 Discriminator network
 CONTI….

Deep Convolution Neural Networks

Architecture of CNN was influenced by the visual cortex structure of human
brain. Essentially, a deep learning algorithm that takes images or spectrograms
and assigns different weights to each image for distinction, and perform any
given task, such as image classification .
 DEEP FAKE
DETECTION
oA criterion was develop to compare different types of CNN
architectures used by several authors for voice recognition,
classification and speech detection.
oCriterion includes:
Input size
The filters and strides
The pooling
The pooling step
The architecture receptive fields
The activation function
The padding
The drop-out regularization
The accuracy
The types of architecture
CONTI….
 EXPERIMENTATION

Dataset
oThe Baidu Silicon Valley AI Lab Dataset is arranged according to
its source in folders
oFolders are split into two:
The original audio
The deep fake audio
oAs per standard practice in machine-learning research, both pre-
processed versions of dataset have been divided in 80/20 split for
training and validation.
PROCESSING
 RESULTS
AND
DISCUSSION
CONTI….
 CONCLUSION

o Mel-Spectrum frequency for visualization provides the most

accurate results from all the optimizers and architectures
explored.
oWhile the results provide a good indication of which
architectures are the best as well as the optimizer, more work
will be carried out to generating more robust and featureful
datasets, such that more effects can be observed.
o The result presented in this study can be leveraged for
multimedia forensic analysis, to guide forensic analysts in
choosing the appropriate approach for deep fake audio analysis.
Thank you for your patience