Protecting Network Quality of Service Against Denial of Service Attacks Douglas S. Reeves S. Felix Wu Fengmin Gong NC STATE UNIVERSITY / MCNC DARPA IA&S Meeting July 20, 2000 NC STATE UNIVERSITY / MCNC New Capabilities... • Different classes of service for users – how much bandwidth – what quality level (delay, loss rates) • Based on trust, need, importance, urgency, .... : Policies! 2 NC STATE UNIVERSITY / MCNC Provided By... • Service provider provisions the resources for the expected demand • User makes request • Network allocates bandwidth amount and quality level, sends response • Network enforces amount / quality 3 NC STATE UNIVERSITY / MCNC …Create New Vulnerabilities! • Each step can be attacked 4 NC STATE UNIVERSITY / MCNC Attack 1: Excessive User Demands • Everyone asks for... – maximum resource amount – premium service • Why not? 5 NC STATE UNIVERSITY / MCNC Our Solution: Resource Pricing • (An example: Telephone Network) 6 NC STATE UNIVERSITY / MCNC Resource Prices Based on Demand • Predicted-load (static) pricing – ex.: provisioning, time-of-day pricing • Auction-based (semi-static) pricing – ex.: bandwidth exchanges, time-slot assignment • Congestion-based (dynamic) pricing – ex.: congestion control • Combined approaches 7 NC STATE UNIVERSITY / MCNC Policy Specification / Enforcement • What determines the price? • How much can each user pay? 8 NC STATE UNIVERSITY / MCNC Provable Fairness • Fairness is the consequence of a policy • Achievable... – Pareto optimal – Weighted max-min fair – Proportional fair – Equal QoS – Maximal aggregate utility – Maximum revenue 9 NC STATE UNIVERSITY / MCNC Properties • Simple, distributed computation • Fast convergence • Low overhead 10 NC STATE UNIVERSITY / MCNC Comparison With Other Approaches • First-come, first-served – “grab resources early and often” • Fixed (absolute) priority – starvation problems • Non-weighted fairness (TCP) – everyone is equal? • Other resource pricing work – static / centralized, restricted fairness 11 NC STATE UNIVERSITY / MCNC Future Work: Implementation • Fall 2000 (management tools: Summer 2001) 12 NC STATE UNIVERSITY / MCNC Fut. Wk.: 3rd Party Authorization • Fall 2000 13 NC STATE UNIVERSITY / MCNC Future Work: Service Class Provisioning • Given predicted demand for each service class... – how much of each service class should network owner provision? – what price charge for each class? • Goals: maximum profit, maximum utility, ...? • Spring 2001 14 NC STATE UNIVERSITY / MCNC Future Work: Protecting the Pricing Mechanism • Vulnerability to attack • Protecting… – RSVP – COPS – SIP – Policy server and databases – Authorization server, user database, billing database • Spring 2002 15 NC STATE UNIVERSITY / MCNC Impact of This Work • Disincentives for "bad" user behavior • Ability to flexibly specify and enforce policies • Efficient (optimal) allocation • Economic incentives for deployment of new services 16 NC STATE UNIVERSITY / MCNC Attack 2: Modify Resource Request / Response Signals • RSVP is the control mechanism for QoS • Routers can "legally" modify these signals • How detect illegal modification? 17 NC STATE UNIVERSITY / MCNC RSVP Attack Examples 18 NC STATE UNIVERSITY / MCNC RSVP Attack Examples 19 NC STATE UNIVERSITY / MCNC RSVP Attack Examples 20 NC STATE UNIVERSITY / MCNC Our Solution: Selective Signing + Auditing • 1. Sign at the end-points the message contents that cannot be changed • 2. Each router audits fields that can be changed – remember values transmitted downstream – compare with values propagated upstream • Has been implemented and tested 21 NC STATE UNIVERSITY / MCNC Comparison With Other Approaches • End-to-end signing of complete message contents – Won't work with changeable contents • Hop-by-hop signing of message contents – Excessive overhead – Does not detect attacks by corrupted routers 22 NC STATE UNIVERSITY / MCNC Future Work • Other attacks – Message dropping – Message insertion – Resource "hoarding" – Summer 2001 • Auditing – Integration with intrusion detection – Fall 2001 23 NC STATE UNIVERSITY / MCNC Impact of This Work • Practical, more effective detection of DoS attacks on control flow 24 NC STATE UNIVERSITY / MCNC Attack 3: TCP Packet Dropping • Congestion causes "normal" packet dropping • Can malicious packet dropping (not due to normal congestion) be detected? – due to corrupted routers – due to "unfriendly" users 25 NC STATE UNIVERSITY / MCNC Our Solution • Build a profile of normal dropping behavior • Compare with observed dropping behavior – statistical techniques – neural net techniques • Experiments: 5 sites in 4 countries over several weeks 26 NC STATE UNIVERSITY / MCNC Effectiveness • Created several types of dropping attacks – random – periodic – re-transmissions only • Measured losses and latency • High detection rate (> 80%), low (1%5%) false positive rate 27 NC STATE UNIVERSITY / MCNC Impact • Attacks will become less obvious; degraded service, not disrupted service • First work on monitoring this type of attack 28 NC STATE UNIVERSITY / MCNC Attack 4: Compromised DiffServ Routers 29 NC STATE UNIVERSITY / MCNC Attack Types • Dropping one data flow to benefit others • Injecting(spoofing, flooding,...) packets to a high priority flow • Remarking packets in a data flow • Delaying packets in a data flow • Compromised ingress, core, or egress routers 30 NC STATE UNIVERSITY / MCNC Approach • Monitor router behavior externally • Monitoring agents externally controlled by intrusion detection system (IDS) – selectively enabled when needed • IDS performs analysis of measurements 31 NC STATE UNIVERSITY / MCNC Monitoring • Granularity is Per-Hop-Behavior (PHB, macroflow) • Metrics: ingress rate, egress rate, drop rate, delay • Passive (packet counting) • Active (packet probing) 32 NC STATE UNIVERSITY / MCNC Status • Attack analysis • Architecture • Testbed, measurements • Future work – Implement passive monitoring, Fall 2000 – Implement active monitoring, Spring 2001 – Implement analysis, Summer 2001 – Integrate with existing intrusion-detection engine, Year 3 33 NC STATE UNIVERSITY / MCNC Impact • First work on detecting and preventing attacks on DiffServ 34 NC STATE UNIVERSITY / MCNC Technology Transfer • Code release (pricing simulator, TCP dropping attack analyzer) • Patent application on pricing with NEC • Collaboration with Nortel on resource authorization • Discussions with Enron, NEC, IETF DiffServ WG 35 NC STATE UNIVERSITY / MCNC General Hicks’ “Hot List” of Needs • Prevent Denial of Service attacks • Automate network bandwidth allocation – reallocate to other, changing priorities 36