Towards Anomaly/Intrusion Detection and Mitigation on High-Speed Networks

Towards Anomaly/Intrusion Detection and Mitigation on High-Speed Networks Yan Gao, Zhichun Li, Manan Sanghi, Yan Chen, Ming-Yang Kao Northwestern Lab for Internet and Security Technology (LIST) Department of Computer Science Northwestern University http://list.cs.northwestern.edu

Outline • Motivation • Architecture of RAIDM • Statistical Sketch-based Anomaly/Intrusion Detection • Design • Evaluation Results • Polymorphic Worm Signature Generation with Provable Attack Resilience • Design • Preliminary Evaluation Results • Conclusions

Desired Features of Intrusion Detection Systems (IDS) • Network-based and scalable to high-speed links • Slammer worm infected 75,000 machines in <10 mins • Flash worm can take less than 1 second to compromise 1M vulnerable machines in the Internet [Staniford04] • Host-based schemes inefficient and user dependent • Have to install IDS on all user machines ! • Existing network IDS unscalable: In a 10Gbps link, each 40-byte packet only has 32ns for processing ! • Aggregated detection over multiple vantage points • Multi-homing, load balancing, policy routing become popular asymmetric routing • Even worse… Per-packet load balancing • Cannot afford to move traffic around

Features of Demanded Intrusion Detection Systems (IDS) • Attack resiliency • Human adversaries are difficult to handle. Attacks tend to fool or DoS the IDS system. • Many IDSs use exact per-flow states, which is vulnerable to DoS attack. • Attack root cause analysis for mitigation • Detection is only the first step. We also need mitigation and prevention. • Overall traffic based approaches can scale to high speed but cannot really help mitigation • We need flow-level information for mitigation • Signature generation for polymorphic worms

Router-based Anomaly/Intrusion Detection and Mitigation System (RAIDM) • Online traffic recording • Design reversible sketch for data streaming computation • Record millions of flows (GB traffic) in a few hundred KB • Online flow-level anomaly/intrusion detection & mitigation • As a first step, detect TCP SYN flooding, horizontal and vertical scans even when mixed • Existing schemes like TRW/AC, CPM will have high false positives • Infer key characteristics of malicious flows for mitigation • Dos attack resiliency • Apply to distributed detection environment RAIDM: First flow-level intrusion detection that can sustain 10s Gbps bandwidth even for worst case traffic of 40-byte packet streams

RAIDM system RAIDM system Internet scan port Internet LAN RAIDM system LAN Internet LAN Switch Switch Splitter Switch Splitter Router Router Switch Switch Router scan port LAN Switch LAN LAN HAIDM system (a) (b) (c) Router-based Anomaly/Intrusion Detection and Mitigation System (RAIDM) • Attach to routers as a black box Attach HRAID black boxes to high-speed routers(a) original configuration, (b) distributed configuration for which each port is monitored separately, (c) aggregate configuration for which a splitter is used to aggregate the traffic from all the ports of a router.

RAIDM Architecture Remote aggregated sketch records Sent out for aggregation Part I Sketch-based monitoring & detection Reversible k-ary sketch monitoring Normal flows Sketch based statistical anomaly detection (SSAD) Local sketch records Streaming packet data Keys of suspicious flows Filtering Keys of normal flows Polymorphic worm detection (Hamsa) Signature-based detection Per-flow monitoring Suspicious flows Part II Per-flow monitoring & detection Network fault diagnosis (DOD) Intrusion or anomaly alarms Modules on the critical path Modules on the non-critical path Data path Control path

Statistical Sketch-based Anomaly/Intrusion Detection (SSAD) • Recording stage: record traffic of each router with different sketches, then transfer and combine them to current reversible sketch. • Detection stage: use Time Series Analysis (Holt-Winter and EWMA) to detect large forecast error as anomalies. • False positive reduction & 2D sketch skipped (lack of time)

Statistical Sketch-based Anomaly/Intrusion Detection (SSAD) • Distributed Detection Environment

… … h1(k) 0 0 1 1 K-1 K-1 1 1 … … hj(k) j j hH(k) … … H H Background: Reversible k-ary Sketch • Array of hash tables: Tj[K] (j = 1, …, H) • Similar to count sketch, counting bloom filter, multi-stage filter, … • Update (k, u): Tj [ hj(k)] += u (for all j)

+ = Background: Reversible k-ary Sketch • Estimate v(S, k): sum of updates for key k • Sketches are linear • Can Combine Sketches • Can aggregate data from different times, locations, and sources • Inference I(S, t): output the heavy keys whose values are larger than threshold t

Detection Algorithm • RS((DIP, Dport), SYN-SYN/ACK) • RS((SIP, DIP), SYN-SYN/ACK) • RS((SIP, Dport), SYN-SYN/ACK)

DoS Resiliency Analysis • Possible DoS attack to existing approaches like TRW and TRW/AC • Source spoofed SYN flooding attack • Source spoofed packets to random location • Attack SSAD is difficult. Possible attack is based on creating collisions in sketches. But… • Reverse engineering of hash functions is difficult • The possibility of finding collisions through exhaustive search is very low • Attacks are limited even with collisions: need the cooperation with comprised internal hosts.

Evaluation • Data Sets • NU traces (239M flows, 1.8TB traffic/day) • Lawrence Berkeley National Laboratory (LBL) Trace (900M flows) • Rest of results based on NU trace evaluation • Scalability and memory usage • Total 9.4MB used for recording hundreds of millions of flows • 3 reversible k-ary sketches, 2 two dimensional sketches and 1 original k-ary sketch • Small # of memory access per packet

Evaluation (cont’d) • Fast • Recording speed for the worst case traffic, all 40B pkts • 16 Gbps on a single FPGA board • 526 Mbps on a Pentium-IV 2.4GHz PC • Detection speed • On site NU experiment covering 1430 minutes: 0.34 sec for one minute on average. (std=0.64 sec) • Accurate Anomaly Detection w/ Reversible Sketch • Compared with detection using complete flow-level logs • Provable probabilistic accuracy guarantees • Even more accurate on real Internet traces

Evaluation (cont’d) • 25 SYN flooding, 936 horizontal scans and 19 vertical scans detected (after sketch-based false positive reduction) • 18 out of 25 SYN flooding verified w/ backscatter • Complete flow-level connection info used for backscatter • Scans verified (all for vscan, top and bottom 10 for hscan) • Unknown scans also found in DShield and other alert reports Bottom 10 horizontal scans Top 10 horizontal scans

Requirements for polymorphic worm signature generation • Network-based • Worm spread at exponential speed, at early stage there are limited worm samples. • Keep up with the network speed ! • Noise tolerant • Most network level flow classifier may suffer false positives. • Attack resilience • Attackers always try to evade the signature generation system • Efficient signature matching • Again, on high-speed networks! • No existing work satisfies these requirements

Hamsa • Content based v.s. behavior based signature • Content based: treat a worm as a byte sequence • Behavior based: the actual dynamics of the worm execution • Fast matching could be a problem for behavior based • We propose Hamsa, a network-based signature generation system to meet the aforementioned requirements • Content based system (token based) • Accuracy comparable to the state-of-the art technique [Polygraph] but much faster • Propose an adversary model and has attack resilience guarantee

Hamsa Design • Sniff traffic from networks • Assembly the packets to flows • Classify traffic based protocol • Filter out known worms • Generate suspicious and normal pools.

Hamsa Design • Iterative signature generation • Extract the tokens form suspicious pool • Identify the same set of tokens in normal pool • Core part: greedy algorithm based on the token information

Model-based Polymorphic Worm Signature Generation Problem & Algorithm The problem No noise: O(n) Noise: NP-Hard The model The algorithm

Evaluation Methodology • Data collection • Normal pool: • 300MB Web traffic for training • 20GB Web traffic and Binary distribution of Linux for evaluation • Suspicious pool: • 10 ~ 500 samples (with noise and different worms) • 5000 samples each worm as false negative testing data set • Three pseudo worms: ATPhttpd, Apache-Knacker, Codered II • Two polymorphic engines from Internet • Simulation settings • Single worm with noise • Test pool size 100 and 200 • Noise ratio 0, 10%, 30%, 50% • Multiple worms with noise • 3 worms together: ATPhttpd, Apache-Knacker, Codered II • Test pool size 100 and 200 • Noise ratio 0, 10% and 25%

Preliminary Evaluation Results • Signature Quality • Signature Generation Speed • 64 ~ 361 times faster than Polygraph • Only use several seconds to at most minutes to generate signatures • Attack Resilience • Provable attack resilience (details omitted) • For token-fit attack, Polygraph fail but Hamsa succeed

Backup Slides

Intrusion Detection and Mitigation

ErrorSketch Sketchmodule Forecastmodule(s) Anomaly detectionmodule (k,u) … Alarms Sketches Reversible Sketch Based Anomaly Detection • Input stream: (key, update) (e.g., SIP, SYN-SYN/ACK) • Summarize input stream using sketches • Build forecast models on top of sketches • Report flows with large forecast errors • Infer the (characteristics) key for mitigation

Reducing False Positives for SYN Flooding Detection • Reasons of false positives of SYN flooding detection • Network/server congestions/failtures • Polluted or outdated DNS entries • Filters to reduce false positives caused by bursty network /server congestions/failures 1. 2. Lifetime > Thresholdlife • Filters to reduce the false positives caused by misconfigurations or related problems • No connection history

Intrusion Classification with Two Dimensional Sketch • We need to distinguish different types of attack to take the most effective mitigation scheme • However, one dimensional information is not enough • Non-spoofed SYN flooding v.s. horizontal scan {SIP,DP} • Bi-modal distribution. • Two dimensional Sketch Structure of 2D sketch Example UPDATE

Intrusion Classification with Two Dimensional Sketch • Two dimensional sketch is accurate • Refer to the paper about accuracy proof • Can archive 99.9956% detection rate by the parameter mentioned in the paper

Automated worm signature generation • Manual signature generation is too slow for fast propagated worms • Automatic signature generation has been proposed [Earlybird][Autograph] • But it is not hard for hackers to apply polymorphism to their worms. • Polymorphic worm signature generation becomes necessary.

Towards Anomaly/Intrusion Detection and Mitigation on High-Speed Networks