210 likes | 349 Views
Evolving Intrusion Detection System for MLDB. Muthukumar Narayanan Final Presentation for CS401 11/22/2004. Overview. Motivation MLDB Intrusion Detection Problem Statement EA Implementation Experiments Results Conclusion Future Work. Motivation. Government Info. Medical Info.
E N D
Evolving Intrusion Detection System for MLDB Muthukumar Narayanan Final Presentation for CS401 11/22/2004
Overview • Motivation • MLDB • Intrusion Detection • Problem Statement • EA Implementation • Experiments • Results • Conclusion • Future Work
Motivation • Government Info. • Medical Info. • Bank Account Info.
MLDB • Multi Layered Database • Several layers of information. • Lowest layer corresponding to the primitive and most secure information. • Higher layers store more general and less secure information extracted form one or more lower layers. • Generalization is based on the concept hierarchies
Intrusion Detection • Anomaly Detection • Detects insider attacks • Uses signature of the normal user activities • Misuse Detection • Detects intrusive activities • Uses signature of the intrusive activities • Types of database intrusions • Inferences, SQL injection, Buffer overflow attack, password attack . . .
Inference Example 1. Select avg(Salary) From Employee_details_noFD; 2. Select count(Salary) From Employee_details_noFD; 3. Select avg(Salary) From Employee_details_noFD Where Name <> "X8” (Result (1)*Result (2)) - (Result (3)*(Result (2)-1))
Problem Statement • Evolve intrusive queries for a given database relation using Genetic Programming approach. • Use them as a rule base for detecting real world intrusions
Representation • SQL query is converted to the corresponding relational algebraic expression. • Select Name from table_1 where Grade=‘A’ or Grade=‘B’; • Relational Algebraic expression is • ПName(σ(Grade=‘A’ or Grade=‘B’)(table_1)) • The tree is represented based on the Relational Algebraic expression
Fitness Evaluation • Secured attribute in the Projection Operation (5-10) • Statistical information excluding one tuple (8-10) • Statistical information based on the attributes involved in functional dependency(8-10) • Using sensitive attributes in the selection operation(1-5) • Using Statistical operation on sensitive data (1-5) • Secured attribute involved in a value constraint(5-10)
Experimental Setup • Relation {Emp_id, Name, Designation, Status, Projects, Deductions, Salary, Net} • Functional Dependency • Designation -> Salary • Value constraints • Salary-Deductions = Net • Protected attributes • Net and Projects
Initial Queries • Select status, salary from MLDB_table where Name = constants; (9) • Select Net, Name, Deductions, Emp_id from MLDB_tab;e where Deductions > constants; (22)
Evolved Queries • Select avg(Net), deductions, designation, avg(salary), Net, count (salary), Projects, avg(deductions), avg(salary) from MLDB_table where (Emp_id <> constants OR Salary <> constant) AND deductions <> constant AND Emp_id <> constant. (104) • Select Net, Net, Net, Net, Net, Net, Net from MLDB_table where . . . (184)
Evolved Queries • Select Salary, Max(Net), Projects, Emp_id, Name, avg(Salary), Net, count(Deductions) form MLDB_table where Name = constant AND Emp_id <> constants; (74) • selectavg(Net), Emp_id, Emp_id from MLDB_table where Designation = constant; (22)
Conclusion & Future Work • Results are NOT complete but satisfactory at this stage • Modified Fitness Evaluation • Allow only LEGAL Queries to evolve • Use of various other large database relations • Use of more Stochastic based parent and survivor selection