A neural network is powerful. Within limits, it supposed to be able to compute any arbitrary function. But since any process can be computed as a function, neural networks can be used for all sorts of things from translating a page written in Cyrillic into English to picking out pictures of domestic cats from a whole menagerie of animal photos. It can name a tune from mere bits of a song or summarize a movie.

Neural networks, however, have one glaring hole. They are a black box.

This is a big problem when the tools of digital forensics are used in court. The United States has an accepted rule regarding the admissibility of expert witnesses. It is called the Daubert Standard. And processes that cannot be explained will be a sore point with both the judge and the jury.

A solution to this problem may be the addition of fuzzy logic to the neural network. Instead of the binary 0 and 1, yes or no, of hard computing, fuzzy logic allows the use of intermediate steps such as all the gradations between 0 and 1.

This new combination gave us fuzzy logic systems with the ability to learn. Instead of the black box that is the pure neural network, the adaptive neuro-fuzzy system (AI with fuzzy logic) allows a view into how the whole thing does what it does because it is governed by a system of If…Then rules.

Adaptive neuro-fuzzy (ANF) is particularly good at modelling systems. Once the system is modelled, its internal rules can be examined and experts can cite the rules that govern its operations. This could lead to new breakthroughs in detecting possible fraud in financial systems, to give one example.

From a paper presented by Andrii Shalaginov from the Proceedings of the Twenty-Sixth International Joint Conference on Artificial Intelligence (IJCAI-17):

We believe that data-driven approach based on Fuzzy Logic may be beneficial for data representation in a Court of Law. In particular, a trade-off between the accuracy of the model and the interpretability can be optimized using Neuro-Fuzzy in Digital Forensics applications. The improved model includes exploratory data analysis for optimal SOM configuration, new fuzzy patches and membership function. Wide range of experiments proved the ability to extract lower amount of fuzzy logic rules and achieve higher classification performance on malware detection and network traffic analysis problems.

Keep reading here.