Deep fake image and video tampering detection by Professor Sabah Jassim

Topological Data Analysis (TDA) and its main tool of Persistent Homology (PH) aims to prevent and detect deep fake image attacks on face biometric systems by using classifying Persistent Homology parameters computed for automatically selected image landmarks. An example of such an attack would be defined by combining two face images of two different people to create a fake ID document for an individual who is not legally entitled to a passport/ID. By doing this, these images will appear genuine but in reality, are fake images and will fool the ID detection systems, allowing them to enter illegally.

Morph effects are amendable to TDA via its main tool of Persistent Homology whereby the Persistent Homology model changes in the topological invariants of nested “rips”, simplicial complexes generated by certain automatically determined landmarks at increasing distance resolutions. The Persistent Homology associated with certain landmarks and compute topological invariants at an increasing sequence of distance thresholds. Persistent Homology barcodes with persistent diagrams shows the changes in the number of connective composition holes and creates a completed graph by filling the gaps.

Basic Machine Learning algorithms applied to the Persistent Homology parameter vectors, were shown to be an effective morph detection tool.  Results of the effectiveness of this approach were demonstrated by testing the scheme with large sets of fake morphed as well as genuine images from Utrecht and London public image databases. The PH-based morph detector performed higher on Utrecht dataset than the London dataset due to the higher diversities in participants’ background.

Comparing the PH parameters extracted from a small set of deep fake videos and original videos, has shown that PH is capable of separating them with high chance of detection. This pilot study has confirmed that Persistent Homology provides a highly effective way of detecting deep fake videos.

In conclusion, TDA is a promising Artificial Intelligence approach to detect image tampering and the detection can be optimised by expanding the list of landmarks, e.g. local derivative pattern codes.

Professors and lecturers of the School of Computing give postgraduate and undergraduate students an insight into their research, showing them range of research projects which have been undertaken. This is an opportunity to broaden their understanding of computing and identify areas of interest for further study. All are encouraged to attend.

Check out the upcoming School of Computing events and seminars.

Find out more about our courses in computing.