An Analysis of the Effectiveness of a Constructive Induction-Based Virus Detection Prototype

Computer viruses remain a tangible threat to the systems on which the Department of Defense increasingly depends. This threat is exacerbated continually, as new viruses are introduced at an alarming rate by the growing collection of connected machines. Unfortunately, current antivirus solutions are ill-equipped to address these issues in the long term. This thesis documents an investigation into the use of constructive induction, a form of machine learning, as a supplemental antivirus technique theoretically capable of detecting previously unknown viruses through generalized decision-making techniques. A group of examples derived from common software applications, utilities, and viruses was tested to evaluate the benefits of utilizing constructive induction in the selection of suitable virus signatures. A prototype virus detection system subcomponent, DRIVER, was developed to conduct the experiments. Due to the feature-rich content of nontrivial example files and DRIVER's ability to assemble decision trees, results showed marginal benefits-compounded with significantly increased computational resource requirements--in the use of constructive induction. Future research, emphasizing a combination of optimization techniques and increasingly realistic test cases, should eventually establish whether constructive induction represents a practical alternative to today's antivirus measures.