Distributed Firewall Policy Validation by Kyle Wheeler
Outline 1. Introduction Justification Requirements 2. Design Approaches Architecture 3. Implementation Requirements Graphing Example Policy Example 4. Conclusions
Security is IMPORTANT Computer-based attacks are increasing Code Red: 2000 hosts/minute (2001) Slammer: 55 million scans/second (2003) Attacks are becoming more damaging CISCO’s IOS code stolen Valve’s HalfLife 2 code stolen Trend Micro says: $13 billion in 2001 $20 billion in 2002 $55 billion in 2003 (source)source
Security is HARD Firewalls Most popular security method Rules can and do become very complex Not only method, however Large networks have: Many different administrators Diverse software Security of large networks requires: Centralized control Uniform software No unified method of verifying security policy implementation For example, The University of Notre Dame network
Rules for the Solution Few Requirements Network-connectivity independent Mostly system-setup independent Cannot require root access Independent of firewall implementations Flexible Testing Out-of-order data collection (some support) Non-uniform distribution of testing nodes Define a testable security policy language
Analysis Approaches Static Vulnerability Analysis Splint Threat Modeling Regression Testing
Static Vulnerability Analysis The Good Avoids logical ambiguity Avoids common loopholes and mistakes Easy to understand The Bad Requires detailed knowledge of the implementation Implementation- specific Does not address system interactions
Threat Modeling The Good Models entire system Views system as an attacker would Determines vulnerability “surface” The Bad Requires full knowledge of all system details
Regression Testing The Good Does not need implementation- specific details Easy to understand The Bad Effectiveness is tied to the completeness of the policy Can miss some vulnerabilities
Data Collection Framework Hierarchical organization Handles complex networks Allows asynchronous operation Wizard Big picture management, handles policy testing setup Manager Organization, Coordination, Retrieval Prober Low-level testing, yes/no answers
Managers & Probers Good Features Subordinate Managers Commands can be any length Key Features Hierarchical Naming Maildir-like communication
Hierarchical Naming Names contain routing information Names are given or assigned Network must be laid out intelligently No auto-discovery Manually connectable Must be a root to the tree (base) Three kinds of sub-names base.m1.m1.p2.t1.t Example, slide 17, 12
Maildir-like Algorithm Benefits No locks: NFS safe No partial-files No new communication server to secure Two-step file creation Create in tmp, then move to new Need unique new name Use pid and random Could use more (inode#, for example) Waiting For Results Requires Polling
Given a complex network… Administrator’s Console Firewall Prober Manager Prober Manager Prober Manager Prober
… Handled Nicely Prober Manager Prober Manager Prober Manager Administrator’s Console Firewall
Or, More Realistically… internet
... Which Can be Organized Wizard & Manager base Prober base.p Prober base.p Prober base.m1.p Prober base.m1.m1.p Prober base.m1.m1.p Manager - base.m1 Prober - base.p Manager - base.m1.m1 Prober - base.m1.p Prober base.p4
The Implementation Requirements: ttcp installed in PATH Binary connection testing bash available, in PATH Written in bash SSH access, without password Security issue Impact can be reduced with careful administration Graphing with Graphviz
Raw Manager Capability Hosts, fully connected: wopr.memoryhole.net iss.cse.nd.edu salinan.cse.nd.edu itisfast.cse.nd.edu Legend: Black line = confirmed connection Dotted line = one side reported connection Red line = one side reported, one side denied
The Wizard Interchangeable element Interprets policy language Generates and spawns tests At least three per assertion Otherwise 50% of all possible Interprets results of tests Must have control of “base” Manager
Example Policy network iss network nd network brk brk -> nd brk -> iss via nd -> brk via nd -> iss via iss -X nd iss -X brk 16
Conclusions Design is feasible Implementation works as expected Being generic is hard Future Work Investigate long-running “continuous” testing Policy language needs further flexibility Speed of testing
Any Questions?