Defending Distributed Systems Against Malicious Intrusions and Network Anomalies Kai Hwang Internet and Grid Computing Laboratory University of Southern California Keynote Presentation at the IEEE International Workshop on Security in Systems and Networks (SSN-2005), held in conjunction with the IEEE International Parallel and Distributed Processing Symposium (IPDPS-2005), Denver, Colorado, April 8, 2005 This presentation is based on research findings by USC GridSec team. Project Web site: http://GridSec.usc.edu, supported by NSF ITR Grant No. 0325409, and contributed by Min Cai, Shanshan Song, Ricky Kwok, Ying Chen, and Hua Liu 1 Presentation Outline: Security/privacy demands in networked or distributed computer systems GridSec NetShield architecture for defending distributed resource sites in Grids, clusters, etc. Internet datamining for collaborative anomaly and intrusion detection system (CAIDS) with traffic episode rule training and analysis Fast containment of internet worm outbreaks and tracking of related DDoS attacks with distributedhashing overlays April 8, 2005, Kai Hwang http://GridSec.usc.edu 2 Security and Privacy Demands in Network and Distributed Systems Trusted resource allocation, sharing, and scheduling Secure communications among resource sites, clusters, and protected download among peer machines Intrusion and anomaly detection, attack repelling, trace back, pushback of attacks, etc Fortification of hardware/software (firewalls, packet filters, VPN gateways, traffic monitors, security overlays, etc. ) Self-defense toolkits/middleware for distributed defense, risk assessment, worm containment, response automation Anonymity, confidentiality, data integrity, fine- grain access control, resolving conflicts in security policies, etc April 8, 2005, Kai Hwang http://GridSec.usc.edu 3 GridSec: A Grid Security ITR Project at USC Site S1 Host 3 VPN Gateway 3 Host Internet 3 Host 2 3 3 Host 2 3 Host 3 1 Site S2 Host 3 VPN Gateway Host Host 3 VPN Gateway 3 Host Steps for automated self-defense at resource site : Step 1: Intrusion detected by host-based firewall /IDS Step 2: All VPN gateways are alerted with the intrusions Step 3: Gateways broadcast response commands to all hosts April 8, 2005, Kai Hwang http://GridSec.usc.edu 4 Site S3 The NetShield Architecture with Distributed Security Enforcement over a DHT Overlay Invoke Response Broadcast Update Prevalence Flood control for DDoS Defense Worm Signature Generation Signature Update Collaborative Alert Correlation WormShield and DDoS defense Misuse Detection Anomaly Detection CAIDS Distributed Intrusion Detection/Response System Intrusion Detection Information Exchange DHT-based Overlay Network Security Policy Implementation Overlay Network for Trust Management Authentication Authorization Delegation Integrity Control Trust Integration/Negotiation Platform Overlay April 8, 2005, Kai Hwang http://GridSec.usc.edu 5 Building Encrypted Tunnels between Grid Resource Sites Through the DHT Overlay The number of encrypted tunnels should grow with O(N) instead of O(N x N), where N is the number of Grid sites Using shortest path, security policy is enforced with minimal VPN tunnels to satisfy special Grid requirements, automatically How to integrate security policies from various private networks through the public network ? How to resolve security policy conflicts among hosts, firewalls, switches, routers, and servers, etc. in a Grid environment ? April 8, 2005, Kai Hwang http://GridSec.usc.edu 6 Trust Integration over a DHT Overlay V Site S3 V Site S2 Site S1 Physical backbone Site S4 DHT Overlay Ring V V V VPN Gateway SeGO Server Trust Vector Trust vector propagation User application and SeGO server negotiation Hosts Cooperating gateways working together to establish VPN tunnels for trust integration April 8, 2005, Kai Hwang http://GridSec.usc.edu 7 USC NetShield Intrusion Defense System for Protecting Local Network of Grid Computing Resources ISP The Internet April 8, 2005, Kai Hwang Network Router The NetShield System Firewall Datamining for Anomaly Intrusion Detection (IDS) Risk Assessment System (RAS) http://GridSec.usc.edu Intrusion Response System (IRS) 8 Victim’s Internal Network Alert Operations performed in local Grid sites and correlated globally Local alert correlation Global alert correlation DHT module Global alert clustering Alert classification Alert merging Alert formatting Alert correlation Alert clusters Local alert clustering Alerts IDS IDS IDS April 8, 2005, Kai Hwang http://GridSec.usc.edu Intrusion reports Alert Assessment Reporting, and Reaction 9 Basic Concept of Internet Episodes Event Type: A, B, C, D, E, F, etc. Event Sequence: e.g., <(E,31),(D,32),(F,33)> Window: Event sequence with a particular width Episode: partially ordered set of events, e.g. whenever A occurs, B will occur soon Frequency of episode: fraction of windows in which episode occurs Frequent episode: set of episodes having a frequency over a particular frequency threshold Frequent episode rules are generated to describe the connection events April 8, 2005, Kai Hwang http://GridSec.usc.edu 10 Frequent Episode Rules (FER) for Characterizing Network Traffic Connections E → D, F ( c, s ) The episode of 3 connection events (E, D, F) = (http, smtp, telnet). On the LHS , we have the earlier event E (http). On the RHS, we have two consequence events D (smtp) and F(telnet); where s is the support probablity and c is the confidence level specified below: (service = http, flag = SF) → (service = smtp, srcbyte = 5000), (service = telnet, flag = SF) (0.8, 0.9) Support probability s = 0.9 and Confidence level c = 0.8 that the episode will take place in a typical traffic stream April 8, 2005, Kai Hwang http://GridSec.usc.edu 11 A Cooperative Anomaly and Intrusion Detection System (CAIDS), built with a Network Intrusion Detection System (NIDS) and an Anomaly Detection System (ADS) operating interactively through automated signature generation Training data from audit normal traffic records Single-connection attacks detected at packet level Audit records from traffic data IDS Known attack signatures from ISD provider April 8, 2005, Kai Hwang Signature Matching Engine Attack Signature Database Unknown or burst attacks New signatures from anomalies detected Episode Rule Database Episode Mining Engine Anomalies detected over multiple connections ADS Signature ADS http://GridSec.usc.edu Generator 12 Internet Datamining for Episode Rule Generation Audit data Feature extraction Connection Records Training phase Attack-free episode rules Episode rule mining Engine Detection phase April 8, 2005, Kai Hwang Normal profile database Rules from real-time traffic Anomaly Detection Engine http://GridSec.usc.edu Alarm Generation 13 Attack Spectrum from MIT Lincoln Lab in 10 Days of Experimentation Attack num bers 20 15 DoS U2R 10 R2L 5 Probe 0 Day1 Day2 Day3 Day4 Day5 Day6 Day7 Day8 Day9 Days April 8, 2005, Kai Hwang http://GridSec.usc.edu 14 Day10 Automated Signature Generation from Frequent Episode Analysis 1. Label relevant connections to associate with an FER. Online traffic episode rules from the datamining engine Episode rules matching the normal FER database ? Yes Episode Frequency exceeding the rule threshold ? No Yes No (Massive attacks) 2 Calculate additional information such as connection count, average and percentage of connections, etc. 3 Select one of the predefined classifiers 4 Use the selected classifier to classify the attack class and find the relevant connections 5 Extract common features in all identified connections, such as the IP addresses, protocol, etc. to form the signature (Stealthy attacks) 2 Check error flags or other useful temporal statistics 3 Extract common features such as IP addresses, protocol, etc. to form the signature Adding new signatures to the Snort database Ignore the normal episode rules from legitimate users (No anomaly detected) April 8, 2005, Kai Hwang http://GridSec.usc.edu 15 Successful Detection Rates of Snort , Anomaly Detection System (ADS), and the Collaborative Anomaly and Intrusion Detection System (CAIDS) April 8, 2005, Kai Hwang http://GridSec.usc.edu 16 False Alarms out of 201 Attacks in CAIDS Triggered by Different Attack Types under Various Scanning Window Sizes Number of False Alarms 18 16 14 R2L DoS Pr obe U2R 12 10 8 6 4 2 0 100 300 500 1000 7200 Window Size (Second) Using larger windows result in more false alarms. Shorter windows in 300 sec or less are better in the sense that shorter episodes will be mined to produce shorter rules, leading to faster rule matching in the anomaly detection process April 8, 2005, Kai Hwang http://GridSec.usc.edu 17 Intrusion Detection Rate (%) Detection Rates of Snort, ADS, and CAIDS under Various Attack Classes 70 60 50 40 SNORT ADS CAI DS 30 20 10 0 DoS U2R R2L PROBE Tot al At t ack Types On the average, the CAIDS (white bars) outperforms the Snort and ADS by 51% and 40%, respectively April 8, 2005, Kai Hwang http://GridSec.usc.edu 18 ROC Curves for 4 Attack Classes on The Simulated CAIDS Intrusion Detection Rate (%) 80 70 60 DoS Pr obe R2L U2R 50 40 30 20 10 0 0 2 4 6 8 10 12 False Alarm Rate (%) April 8, 2005, Kai Hwang http://GridSec.usc.edu 19 Intrusion Detection Rate (%) ROC Performance of Three Intrusion Detection Systems 80 70 60 50 40 30 CAI DS Snor t ADS 20 10 0 0 2 4 6 8 12 10 False Alarm Rate (%) April 8, 2005, Kai Hwang http://GridSec.usc.edu 20 Internet Worm and Flood Control: A DHT-based WormShield overlay network is under development at USC. Fast worm signature generation and fast dissemination through both local and global address dispersion Automated tracking of DDoS attack-transit routers to cut off malicious packet flows for dynamic DDoS flood control April 8, 2005, Kai Hwang http://GridSec.usc.edu 21 The WormShield Built with a DHT-based Overlay with Six Worm Monitors Local Table: Chord ID 76 112 55 215 Global Table: Tp 10,Tc 20 Chord ID Identified Worm Signature! 215 180 ... Content Block s4 s5 Global Address Prevelance Dispersion 5+6=11 18 4+8=12 22 Tl 3 Content Block s1 s2 s3 s4 0/256 Local Prevelance 1 4 2 5 (src, dest) Addresses S1(A), D1(A) S2(A), D2(A) S3(A), D3(A) S4(A),D4(A) Site A Site B Site E Local Table: Chord ID 180 nce } vela (A) Pre A), D4 a te ( Upd , 5, S4 4 5, s {21 192 U {21 pdate 5, s P 4, 5 revela , S4 n (A ), c e D4( A )} Site F Site C Tl 3 Local Table: Content Local (src, dest) Block Prevelance Addresses S5(D),D5(D) s5 7 Site D April 8, 2005, Kai Hwang 64 Tl 3 Content Local Chord ID Block Prevelance 215 s4 6 180 s5 4 128 http://GridSec.usc.edu 22 (src, dest) Addresses S4(C),D4(C) S5(C),D5(C) The WormShield Signature Generation Process Monitored DMZ Traffic Rabin Fingerprinting G lo A bd ad lCr e Chord ID s Update P(j), C(j) o ns te DID(j) n is t p Pe re rs v No a io le n nT aYes c eb P(j) > Tp & le && C(j) > T Content Block j L o c a Local lC Chord ID Content Block Prevelance No o n ID(j) j L(i, j) L(i, j)> Tl te n Update L(i,j) t P re Yes v a leL c no c c ea Content l SRC IP DEST IP T Block No aA d bd j S(i, j) D(i, j) |S(i,j)|+|D(i,j)|> Ts ler Update e s S(i,j), D(i, j) s D Yes is p e rs io n T Send updates Process updates a b for P(j) and C(j) to for P(j) and C(j) from le Other monitor root(j) other monitors WormShield Monitors http://GridSec.usc.edu Address Dispersion C(j) Report j as suspected worm Disseminate suspected worm signature j to WormShield network Chord Protocol April 8, 2005, Kai Hwang Global Prevelance P( j) 23 Signature Detection in Worm Spreading and the Growth of Infected hosts for Simulated CodeRed Worms on a Internet Configuration of 105,246 Edge networks in 11,342 Autonomous Systems Containing 338,652 Vulnerable Hosts April 8, 2005, Kai Hwang http://GridSec.usc.edu 24 Effects of Local Prevalence Threshold Worm spreading and the growth of infected hosts April 8, 2005, Kai Hwang http://GridSec.usc.edu 25 Effects of Global Address Prevalence on Worm Spreading and the Growth of Infected Hosts April 8, 2005, Kai Hwang http://GridSec.usc.edu 26 Number of infected hosts at detection time Reduction of Infected Hosts by Independent vs. Collaborative Monitoring over the Edge Networks Average of independent monitors Best of independent monitors Collaborative monitors in WormShield 300000 250000 200000 150000 100000 50000 0 61(0.1%) 612(1%) 6121(10%) 30608(50%) Number of edge networks monitored April 8, 2005, Kai Hwang http://GridSec.usc.edu 27 LogLog Cardinality Summary Ingress Router Identified as an ATR Tracking and Flood Control Attack Flows LogLog Cardinality Summary Legitimate Flow Attack Flows Legitimate Flow Packet/Flow Counting for Tracking Attack-Transit Routers (ATRs) Ingress Router Legitimate Flows LogLog Cardinality Summary Ingress Router Identified as an ATR LogLog Cardinality Summary Tracking and Flood Control LogLog Cardinality Summary LogLog Cardinality Summary Packet-level Traffic Matrix A Last Hop Router Flow-level Traffic Matrix B Victim April 8, 2005, Kai Hwang http://GridSec.usc.edu 28 False Positive Rate of Identified ATRs The false positive of attack-transit routers(%) 100 70 % Percentile 80 % Percentile 90 % Percentile 80 60 40 20 0 0.0 0.2 0.4 0.6 0.8 1.0 The ratio of legitimate trafic to attack traffic April 8, 2005, Kai Hwang http://GridSec.usc.edu 29 Other Hot Security Research Areas: Efficient and enforceable trust models are very much in demand for networked and distributed systems: PKI services, VPN tunneling, trust negotiation, security overlays, reputation system etc. Large-scale security benchmark experiments in open Internet environments are infeasible. The NSF/HSD DETER testbed should be fully used in performing such experiments to establish sustainable cybertrust over all edge networks. Internet datamining for security control and for the guarantee of Quality-of-Service in real-life network applications – Interoperability between wired and wireless networks is a wide-open area for further research. April 8, 2005, Kai Hwang http://GridSec.usc.edu 30 Final Remarks The NetShield built with DHT-based security overlay networks support distributed intrusion and anomaly detection, alert correlation, collaborative worm containment, and flooding attack suppression. The CAIDS can cope with both known and unknown network attacks, secure many cluster/Grid/P2P operations in using common Internet services: telnet, http, ftp, Email, SMTP, authentication, etc. Automated virus or worm signature generation plays a vital role to monitory network epidemic outbreaks and to give early warning of large-scale system intrusions, network anomalies, and DDoS flood attacks. Extensive benchmark experiments on the DETER test bed will prove the effectiveness. April 8, 2005, Kai Hwang http://GridSec.usc.edu 31 Recent Related Papers: 1. M. Cai, K. Hwang, Y. K. Kwok, Y. Chen, and S. S. Song, “Fast Containment of Internet Worms and Tracking of DDoS Attacks with Distributed-Hashing Overlays”, IEEE Security and Privacy, accepted to appear Nov/Dec. 2005. 2. K. Hwang, Y. Kwok, S. Song, M. Cai, R. Zhou, Yu. Chen, Ying. Chen, and X. Lou, “GridSec: Trusted Grid Computing with Security Binding and Self-Defense against Network Worms and DDoS Attacks”, International Workshop on Grid Computing Security and Resource Management (GSRM’05), in conjunction with ICCS 2005, Atlanta, May 22-25, 2005. 3. M. Qin and K. Hwang, “Frequent Episode Rules for Internet Traffic Analysis and Anomaly Detection”, IEEE Network Computing and Application Symp. (NCA-2004), Cambridge, MA. August 31, 2004 4. K. Hwang, Y. Chen and H. Liu, “ Defending Distributed Computing Systems from Malicious Intrusions and Network Anomalies”, IEEE Workshop on Security in Systems and Networks (SSN’05), in conjunction with IEEE IPDPS 2005, Denver, April 8, 2005. April 8, 2005, Kai Hwang http://GridSec.usc.edu 32