Deep Fakes As part of the project we explored the various stages innvolved in the genrating deep fakes.
The landascape of the DeepFake research can be roughly divided into three broad categories:
We also observed a classification based on the kind of features that researchers used to enable efficient detecction of deep fakes. Some of them have been listed below.
The first category of DeepFake detection methods are data-driven, which directly employ various types of DNNs trained on real and DeepFake videos, not relying on any specific artifact.
The second category of DeepFake detection algorithms use signal level artifacts introduced during the synthesis process such as those described in the Introduction.
Third category is based on inconsistencies exhibited by the physical/physiological aspects in the DeepFake videos
The segregation can also be observed in terms of the computational approached employed.
-
Deep Learning Approaches
-
Conventional Approaches
With the fast improvement of DeepFake finders, specialists begin focusing on plan techniques to dodge the phony faces being distinguished. In particular, given a genuine or phony face, avoidance strategies map it to another one that can't be accurately arranged by the cutting edge DeepFake identifiers, stowing away the phony appearances from being found.
We tried different approaches and tried looking at the current existing solutions. We have summaried all our insights and prepared a review under the reports section.
The Lit Review section of thre repo consists of some of the papers that were "must reads" or State of the Art in the domain.