6.5 Using Neural Networks

6.5 Using Neural Networks

To construct a model for detecting perpetrators using an MLP neural network, samples of criminal observations and their attributes are required. It is through the training of these samples that connections among the various layers in the network can be made to converge. This enables a model to be created for recognizing future perpetrators, so that whenever the features of a criminal are input, an alert output is produced, since a network will associate and recognize a profile of similarity.

The neural network is then said to have learned to recognize the features of criminal perpetrators. Once training is complete, the neural network uses these learned patterns to predict the probability that a new individual will exhibit the modeled transaction patterns. Moreover, the trained network is able to recognize new perpetrators that are similar, but not identical to, those used in its training sessions.

This application of training and modeling for profiling and prediction to investigative data mining can be applied in criminal investigations, fraud detection, internal corporate investigations, cybercrimes, system intrusion detection, criminal profiling, and criminal analysis and prevention. The challenge, as some of the case studies will illustrate, is in the encoding of the data. Most law enforcement departments and government agencies have not adequately digitized their criminal records into a structured and uniform manner. Many still use free-form text narrative entries in documenting their criminal cases; this impedes the use of neural-network techniques. The top law enforcement agency in the United States, the FBI, has made great strides in gathering criminal data for its investigations and for reports; however, by its own admission, it has been remiss in its analysis of this data.

If a standardized system of recording crimes can be developed, data mining techniques can begin to be deployed in order for crime to be analyzed, modeled, and to an extent predicted. For this to happen, effort on the parts of the local, state, and federal law enforcement agencies must begin to structure in a machine-readable format how crimes get reported and documented. Once the crime data is standardized and warehoused, it becomes a much easier task to analyze it to extract predictive models from it.

6.6 Why Use Neural Networks?

Neural networks have been demonstrated to be very effective in dealing with noisy input data, such as handwriting and speech recognition and various forms of image processing, which are very difficult to process with the rigid reasoning techniques of statistical systems. The pattern-recognition ability of neural networks has proven extremely effective in predicting and recognizing patterns of consumer behavior and can be used similarly to detect criminal activities. Some neural networks, in conjunction with other technologies, are being used for retina, thumb, and facial recognition applications; they are the "reasoning engines" to these proprietary identification systems.

Neural networks provide solutions to a variety of classification problems, such as speech, character, and signal recognition, as well as functional prediction and system modeling, where the physical processes are not understood or are highly complex. So far, neural networks have not been widely applied to profiling criminals, aside from detecting fraud. However, crime and terrorism in our time cannot exist without generating digital trails and transactions, all of which are subject to scrutiny and evaluation by neural network models. In fact, after 9/11, there were calls in the media for development of this very type of analysis to be accelerated by the government and private industry to foster homeland defense.

The advantage of neural networks lies in their ability to deal with samples of input data and learn quickly from these training sessions. They are often good at solving problems that are too complex for conventional technologies but that a human is able to recognize and learn. However, rather than taking years to train a good fraud specialist to recognize when a crime is being committed, a neural network can be trained in a few minutes, if examples of fraudulent observations are available to create a model. From that point, when a red flag or possible criminal scenario is raised, expert human analysis can be implemented to assess the credibility of the program's alert.

Another distinct advantage of neural networks is that they can be used to process and score hundreds of thousands of records or transactions in milliseconds. Once trained, a neural network tool can generate code that can be used in real-time productions systems. Neural networks are highly accurate, portable, and fast, which when coupled with an investigator's knowledge and intuition can be a hybrid deterrence to future threats from criminals and terrorists.