Hybrid Intelligent Systems for Detecting Network Anomalies Lane Thames ECE 8833 Intelligent Systems
Outline Introduce Preliminary Information about computer attacks and computer networking Present the Implementation details and test results Discuss my future work of incorporating intelligent systems into my network security research
Project Goals Develop a hybrid system that uses Bayesian Learning in conjunction with the Self-Organizing Map Analyze the performance of the various systems: Host-Network based features, Network only based features, Host- Network-SOM based features, and Network-SOM based features
Data Sets UCI Knowledge Discovery in Databases (KDD) KDD CUP 1999 for Intrusion Detection Database
Tool Boxes BN Power Constructor NeticaJ Java based Bayesian Learning Library
Common Types of Attacks Buffer Overflow Attacks Redirects program control flow which causes the computer to execute carefully injected malicious code Redirects program control flow which causes the computer to execute carefully injected malicious code Code can be crafted to elevate the privileges of a user by obtaining super user privileges Code can be crafted to elevate the privileges of a user by obtaining super user privileges
Buffer Overflow
Buffer Overflow-Stack Image Overflow buf with *str so that the Return Address (RA) is overwritten If carefully designed, the RA is overwritten with the address of the injected code (contained in the *str input—shell code) buf SFP Return Address * str Rest of Stack
Buffer Overflow After running the program we get the infamous Microsoft alert In Linux you get “Segmentation Fault”
Buffer Overflow—Exception Info
Buffer Overflow—Stack Trace
Common Types of Attacks Denial of Service (DoS) Exhaust a computer’s resources: TCP SYN flooding attack Exhaust a computer’s resources: TCP SYN flooding attack Consume a computer’s available networking bandwidth: ICMP Smurf Attack Consume a computer’s available networking bandwidth: ICMP Smurf Attack
TCP SYN Flooding Attack
ICMP Smurf Attack Victim Subnet Slaves Master
TCP/IP Layered Architecture Application Layer:(HTTP, SMTP, FTP) Transport Layer:(TCP,UDP) Network Layer:(IP,ICMP,IGMP) Link Layer:(Ethernet, PPP)
TCP/IP Encapsulation Link HeaderNet. HeaderTrans. HeaderApp HeaderApp DataLink Trailer
TCP Header Checksum Dst Port Addr Sequence Number Acknowledgment Number HLEN|Resv|U|A|P|R|S|FWindow Size SRC Port Addr Urgent Pointer Options and Padding
Implementation 2 Types of Bayesian Structures Used Network / Host / SOM Based Features Network / Host / SOM Based Features Network / SOM Based Features Network / SOM Based Features
SOM Details Original SOM for project 1: Time series of 200 connections to an isolated web server Time series of 200 connections to an isolated web server Extract port numbers from TCP Header Extract port numbers from TCP Header SOM Weight vector was a length 200 vector representing various types of destination port number sequences (after training) SOM Weight vector was a length 200 vector representing various types of destination port number sequences (after training)
SOM Details Hybrid System: the SOM was a vector of length 3 and contains the values of the TCP destination port number, the TCP flag value, and the global flag error rate The vector represents one connection record (not a time series of connections) TCP flags: 6 bits (U,A,P,R,S,F) and 2^6=64 possible combinations and not all values are valid, i.e. never have an S and F set simultaneously
Hybrid System Architecture Hybrid System Architecture Init. Train. Data SOM Training Modified Data Struct. Developer Struct. FileProcessed Data Bayesian Trainer Bayesian/SOM Classifier Test Data IDS Classification File (Test Results)
Modified Data Example protocolserviceflagsrcBdstBcntSOMoutserrratererrratetypeAtck tcphttpSF normal. tcphttpSF normal. icmpecr_iSF smurf. icmpecr_iSF smurf. tcpprivateS neptune. tcpprivateS neptune.
Host/Network/SOM Structure
Host/Network/SOM Test Results 65,505 Total Test Cases 65,238 Correctly Classified 99.59% Classification Accuracy
Network/SOM Structure
Network/SOM Test Results 63,297 Total Cases 62,871 Correctly Classified 99.33% Classification Accuracy
Attack Probabilities for a single flow
IDS Output for 30,000 Flows
Table of Results H/NH/N/SNN/S TotalCases CorrectlyClassified % Accuracy 99.26%99.59%96.27%99.33%
Future Work Currently doing research in Network Security NSF Funded project: 3 GT Professors 3 GT Professors 3 GT GRAs 3 GT GRAs 3 Year project 3 Year project
Future Work Currently Developing a “Honey Net” Honey Net: A network consisting of computers and various networking gear that you “WANT” to be hacked!
Future Work Goal: Monitor hacker activities in order to build stronger defenses Goal: Incorporate some of the Intelligent system concepts within the Honey Net to assist in processing the large volumes of data that will be collected (via network sniffers, traffic monitors, host-based software such as tripwire, libpcap programs, etc)