Multiple Classifier Systems

The main scientific interest is in Multiple Classifier Systems (MCSs). MCSs are based on the combination of the outputs of an ensemble of different classifiers. Classification accuracy improvements are possible provided that a suitable combination function is designed and the ensemble of classifiers make different errors.

We are currently investigating two problems related to the design of MCSs:

In order to assess the state of the art of the theory and the applications of multiple classifier systems and related approaches five editions of an International Workshop has been held in Cagliari (June 2000, June 2002, and June 2004) and in the UK (July 2001, and June 2003).

For more information, please visit the Workshop web site or contact me.

Papers available at the web site of the Pattern Recognition and Application Group

Pattern Recognition for Intrusion Detection in Computer Networks

Computer security is now becoming a major concern of modern society as a large fraction of information flows through computer networks. Standard protection mechanisms such as user authentication, service control, and traffic filtering cannot guarantee from the risk of computer attacks. The main reason of the weakness of computer networks lies in the great variability of network traffic, and in the so-called “bugs” always contained in system and application software. As a consequence, it is extremely difficult to design rules apt to selectively block intruders’ traffic while allowing legitimate traffic. To design more flexible systems, a number of research papers recently proposed approaches to intrusion detection based on pattern recognition techniques. The pattern recognition approach is expected to help in extracting complex decision rules, that can hardly be implemented by human experts through rule-based systems. Results presented in the literature clearly show the potential of the pattern recognition approach as well as its drawbacks. In fact, while pattern recognition approaches can detect intrusions for which no specific training data were available, they often produce a large number of false alarms, as legitimate traffic can be classified as being intrusive. The challenges posed by this novel pattern recognition application involve all the design phases of a pattern recognition system, i.e., data collection, feature extraction and selection, classifier design, and performance evaluation..

Additional information available at the web site of the Pattern Recognition and Appliccation Group

Learning in Content Based Retrieval in Image Database

We are currently addressing the problem of the refinement of queries in Content Based Retrieval in Image Database. Many techniques and algorithms that perform a query by content in image database are currently available. Usually they present to the user a ranked set of images that are similar to query image. However they do not incorpate user's feedback on the quality of results, i.e. the subset of the output images that the user does not consider similar to the original query. The aim of our research is to develop a method to incorparate users's feedback in the query process that allows a refinement of the query process.

Papers available at the web site of the Pattern Recognition and Application Group