Abstract:
"
With the increase of computational power and large gathering of data, Deep Learning has
become an important part of today’s business and industrial tasks to the extent of
outperforming and replacing humans with machines. With that said, the question to be
asked is can so much trust be put on Deep Learning? For example, consider autonomously
driving cars, if the Deep Learning algorithm equipped fails identify obstacles it would end
up in a disaster. Considering these risks, the research community has focused on
investigating the threat of associated to Deep Learning.
In the recent past, it was identified that deep learning-based models can be fooled by simply
adding a noise to the image. Following that research investigated on approaches for crafting
this type of noise. Hence, raised the domain of Adversarial Attacks. Various researchers
put forwards ways of crafting these attacks and some of they were able to nullify the
accuracy of Deep Learning models. Further into research domain it was shown that this
risk also exists in the Deep Learning models which was used in the real-world application.
Following the growth of the domain Adversarial Attacks the research community
investigated on the countermeasures and was successful to increase the robustness of Deep
Learning against the adversarial attacks. But most of the highly robust model consumes
high computational resources and some of they are practically not deployable. In the other
hand, various research has shown that model compression can reduce the complexity by
maintaining the accuracy and performance of the model. Linking these dots, the author
hypothesis that effective using model compression will reduces the resource consumption
without affecting the robustness that much.
In this research the author was able to successfully implement SuperDef by apply
Knowledge Distillation on RCAN model for single image super resolution task and use
that as defense against adversarial attacks. The benchmark results of SuperDef outperforms
the previous implementation of defense which uses Image Super Resolution in terms of
performance and robustness. The protype of the SuperDef includes web application to
show the results of the defense against various adversarial attacks.
"
