After the Pan Duo that opened Deepfakes in some developer presses demon case, the artificial intelligence of mix the false with the genuine forges photograph technology, had initiated the wide debate of all circles. Good news is, california university river bank cent school Amit K. Roy-Chowdhury teachs group of guided a research, had opened the deepness nerve network that gave out a high level, can analyse a piece of photograph whether by DeepFakes is forged.
(graph from: California university river bank cent school, viaNew Atlas)
During research, the group is cast to their deepness nerve network fed distort in great quantities and be not distort photograph data set, make corresponding computer number to its.
In the picture that is distorted oneself, the group highlights the border that showed the object that is added in camera lens by digitlization to resemble element. There already was researcher to confirm before this, forge a photograph not common object is attrib border, can actualer object more flowing or differ somewhat.
Although person eye is differ,can detect quite surely these difference, but the computer can is opposite unusual the examination that has one by one like element, input the data set of mark deepness nerve network next.
The inattentive algorithm building a model that this is head of an imitate person, aim to identify the pattern in primitive data. Be based on this, network of this deepness nerve adds institutional identifying number the image that go in to indicate border.
Show subsequently when it come from data set when the exterior, picture that had not seen before this, can discern greatly a photograph that forges through Deepfakes.
What need points out is, although this covers a system to apply to static photograph only at present, but the group is trying hard to differentiate what its apply at video. After all the essence of video, it is the successive picture of one frame frame.
Of course, after devoted and actual try out, network of this deepness nerve does not do the identifying of 100% possibly also forever accuracy rate, training data to reach after all had been marked of purpose by researcher.
About the detail of this research, had been published in close sunrise edition " IEEE Transactions On Image Processing " on video processing periodical. Original title is:
" Decoder Architecture For Detection Of Image Forgeries of – of Hybrid LSTM And Encoder "
