Computer Account Hijacking Detection Using a Neural Network Nick Pongratz Math 340 Neural Networks - Example Simple Network - [!] graphic taken from http://blizzard.gis.uiuc.edu/htmldocs/Neural/neural.html Neural Networks - Backpropagation - [!] graphic taken from http://blizzard.gis.uiuc.edu/htmldocs/Neural/neural.html Computer Security Introduction • General computer use is skyrocketing. • Growing reliance on networks. • Greater need to “keep the bad guys out.” Computer Security Introduction • Reactive Security • Proactive Security Computer Security Introduction - Reactive Security - • Break-in already occurred or is occurring. • Minimize/repair damage already done. • Patch the system against further similar attacks. Computer Security Introduction - Reactive Security - • Current applications: Most virus scanners Misuse detection Most Intrusion Detection Systems Computer Security Introduction - Proactive Security - • • • • • Strong passwords and correct permissions. Secure software and operating systems. Find system insecurities before bad guys do. Physical security. Self-adapting, smart systems. Computer Security Introduction - Proactive Security - • Current applications: Self-assessment Some virus scanners – heuristics Anomaly detection Intrusion Detection Systems - General Info - • • • • Most are reactive. Detect strange behavior. Analyze user I/O, network I/O, processes. Look for misuse and anomalies. Intrusion Detection Systems - Misuse Detection - • Compare activity with “signatures” of known attacks. • Signatures typically hand-coded. • Good for known attacks • Bad for previously unknown attacks Intrusion Detection Systems - Anomaly Detection - • • • • • Compare activity with typical activity “Fingerprints” Adaptive Good for detecting unusual behavior. Not great for realtime monitoring. MY PROJECT: Neural Network Anomaly Detection System Neural Network Anomaly Detection System • • • • • Currently analyses user behavior Checks against fingerprints Extendable Adaptive Semi-hybrid: Mostly reactive, has proactive elements Neural Network Anomaly Detection System - Neural Net Technical Details - • Currently implemented in MATLAB. • Object-oriented. • Uses a feedforward backpropagation neural network. • Input: vector of command-use frequency. • Output: vector of true/false guesses of the corresponding users. Neural Network Anomaly Detection System - System Details - 1. Sysadmin runs logs through trained network. 2. System reports the status of the results. 3. Admin (or an automation system) acts on report. Neural Network Anomaly Detection System - Pros and Cons - • Pros: Accurate Extendable Adjusts • Cons: After-the-fact (not realtime) Training data MUST be legitimate Training can take a while One part of complete security system Neural Network Anomaly Detection System - Future Directions - • • • • • Extend to network communication. Extend to running processes. Include progression information in training. Realtime (?) Automatic response automation (?) Any Questions, Comments, Protests, a Summer Job For Me? Thank You! Nick Pongratz njpongratz@students.wisc.edu http://www.cs.wisc.edu/~nicholau/