Engineering a Distributed Intrusion Tolerant Database System Using COTS Components Peng Liu Lab. for Info. and Sys. Security http://www.liss.ifsm.umbc.edu/ University of Maryland Baltimore County July 2001 1 Motivation reject Unauthorized transactions Authorized but malicious transactions database damage Reference Monitor Authorized but malicious transactions can damage a database to useless ! 2 Existing Practice: Database Assurance •Authentication and access control cannot prevent all attacks •Integrity constraints are weak at prohibiting plausible but incorrect data •Concurrency control and recovery mechanisms cannot distinguish legitimate transactions from malicious ones •Automatic replication facilities and active database triggers can serve to spread the damage network 3 The Context From scratch Application or Transaction Level Our focus DBMS Level Trusted DBMS which protects data on un-trusted storage using signatures OS Level Trusted OS or trusted DBMS loader Hardware Using COTS components Our focus 1. Storage jamming 2. Signed hashes Tamper resistant processing environment Multi-layer Database Survivability 4 Technical Objectives Engineering using COTS components a database system that can survive authorized but malicious transactions •Practical Database Intrusion Tolerance –Our approach: using COTS DBMS as main building blocks •Cost effective Database Intrusion Tolerance –Our approach: multi-layer defense, cost-effectiveness-performance analysis •Comprehensive Database Intrusion Tolerance –Our approach: transaction-level intrusion detection, isolation & masking, damage confinement, assessment, and repair •Adaptive Database Intrusion Tolerance –Our approach: self-stabilization by adaptation 5 Assumptions & Policies •What attacks are you considered? –All and only the attacks through malicious transactions •What assumptions are you making? –The proposed ITS facilities are trusted –The COTS DBMS executes transactions correctly •What policies can your project enforce? –New transactions execution control policy, i.e., stop, continue, or reduce speed –Damage confinement policy, i.e., single-phase or multi-phase –Intrusion detection policy, i.e., suspicion levels for malicious and suspicious trans –Attack isolation and masking policy –Self-stabilization control policy, i.e., the minimum integrity level –etc. 6 Expected major achievements • A cost-effective intrusion tolerant database system prototype • A family of innovative database intrusion tolerance techniques – Transaction-level intrusion detection – Intrusion isolation and masking – Multi-phase damage confinement – On the fly damage assessment and repair (implementation) – Adaptive database intrusion tolerance – Optimized load balance among ITS facilities • Identification and study of such ITS properties as adaptability, stability, and sensitivity 7 Our Approach 8 Installation 9 Scheme 1: preliminary intrusion tolerance User Transactions Damage Confinement Mediator (Policy Enforcement) Repair Transactions Intrusion detector Proofs COTS DBMS Damage Repairer Proof collector Damage Assessor 10 Transaction-Level Intrusion Detection •Our goal: applying existing intrusion detection techniques to identifying malicious transactions •Key issues: –semantics-based intrusion detection –proof collection –using the detector as a protection tool –coupling OS-level and transaction-level intrusion detection SSN Start Date Salary 900000001 01/01/97 $58,000 900000001 01/01/98 $60,000 900000001 01/01/99 $62,000 900000001 01/01/00 $82,000 11 Application-Aware Intrusion Detection Intrusion Detector Rule Registration rule base trails Rule Definition Function DLL Application Semantics Every existing ID algorithm can be specified by a rule Rules capture application semantics Active malicious transaction will be rolled back 12 Damage Assessment and Repair A history B1 G2 time The database G3 B1: read(x,z); write(x) G2: read(z); write(z) G3: read(x,y); write(y) x y z B1 Read-from G2 G3 A repair A dependency graph Undo B1 & G3 Our goal: implementation and evaluation 13 New Progress of Scheme 1 •Since Feb 2001 – The intrusion detector prototype is implemented (using ad-hoc rules) – Scheme 1 was demonstrated on DISCEX II in June – A new testing application is developed •Till now – Scheme 1 is implemented (except the damage confinement part) – The prototype has around 20,000 lines of multi-threaded C++ code, running on top of a NT LAN and an Oracle server – The prototype proxies every user transaction, collects the trails of transactions, detects bad transactions, rolls back active bad transactions, locates and repairs the damage (caused by identified bad transactions), all on-the-fly – The prototype (except the intrusion detector) is tested and evaluated 14 Scheme 1: Publications 1. Peng Liu, Xu Hao, "Efficient Damage Assessment and Repair in Resilient Distributed Database Systems", Proc. 15th IFIP WG 11.3 Working Conference on Database and Application Security, July 15-18, 2001, Ontario, Canada. 2. P. Luenam, P. Liu, "ODAM: An On-the-fly Damage Assessment and Repair System for Commercial Database Applications", Proc. 15th IFIP WG 11.3 Working Conference on Database and Application Security, July 15-18, 2001, Ontario, Canada. 3. S. Ingsriswang, P. Liu, "AAID: An Application Aware Transaction-Level Database Intrusion Detection System", Technical Report, Dept. of Information Systems, UMBC, 2001. 15 A Limitation of Scheme 1 •The purpose of confinement is to prevent damage from spreading •The delay of damage assessment can cause ineffective confinement! User SQL Commands Damage Confinement Mediator (Policy Enforcement) B1’s write sets G2’s write sets Repair SQL Commands Intrusion detector B1 Proofs Proof collector B1 G4 Damage Repairer G2 Damage Assessor The database 16 Scheme 2: multi-phase confinement User SQL Commands Damage Confinement Phase 1 Later phases Mediator (Policy Enforcement) Repair SQL Commands Intrusion detector Proofs COTS DBMS Damage Repairer Proof collector Damage Assessor 17 Multi-Phase Confinement: An example To be confined: all data objects updated after time 5 except the data objects updated by G3 User SQL Commands Damage Confinement G3’s write set is clean Mediator (Policy Enforcement) B1 Repair SQL Commands Intrusion detector B1 Proofs Proof collector Damage Assessor B1[5] G4[15] Damage Repairer G2[7] G3[9] The database 18 Damage Confinement: A Spectrum Maximum damage leakage Zero damage leakage Minimum computing resources Maximum computing resources Minimum integrity Maximum integrity Maximum availability One-phase Minimum availability Approximate multi-phase Multi-phase 19 New Progress of Scheme 2 •Since Feb 2001 – The damage confinement subsystem is 70% designed and 70% implemented •Till now – The multi-phase damage confinement technique is developed P. Liu, S. Jajodia, “Multi-Phase Damage Confinement in Database Systems for Intrusion Tolerence”, Proc. 14th IEEE Computer Security Foundations Workshop, June 1113, 2001, Nova Scotia, Canada 20 A Limitation of Scheme 2 •For accuracy, intrusions can be detected with a significant delay •The delay can cause serious damage when an intrusion is detected •Quicker detection can decrease the amount of damage, but could mistake many legitimate transactions for malicious, and cause denial-of-service An user’s history Attack enforced t1 t2 Attack detected The database •Our goal: decreasing the amount of damage without losing detection accuracy and denial-of-service 21 Scheme 3: Isolation User SQL Commands Damage Confinement Suspicious trans. Mediator (Policy Enforcement) Intrusion detector Main database Isolating ... Isolating engine 1 engine n Damage Repairer read Damage Assessor merge 22 New Progress of Scheme 3 • Since Feb 2001 • We have developed a SQL statement rewriting technique to enforce isolation in COTS DBMS • The isolation subsystem is 100% designed • The implementation of the isolation subsystem has started P. Liu, “DAIS: A Real-Time Data Attack Isolation System for Commercial Database Applications”, submitted to ACSAC 2001. 23 A Limitation of Scheme 3 •To reduce cost, very few users (i.e., one) can be isolated within a single engine •However, to avoid causing damage on the main database, the number of suspicious transactions can be large •Hence, isolating every suspicious transaction can be too expensive! •Our solution •Treating very suspicious and less suspicious users differently •Isolating very suspicious users •Masking less suspicious users •Advantage: better cost-effectiveness 24 Scheme 4: Masking User SQL Commands Damage Confinement Mediator (Policy Enforcement) Less suspicious trans. Very suspicious trans. Intrusion detector Damage Assessor Damage Repairer Masking engine 1 Main DB Isolating engine 1 ... Isolating engine n ... Masking engine n read merge 25 Intrusion Masking: An Example Ui : Ti1 Three less suspicious users: Main history Uj : T j 1 Uk : T k 1 Masking history 1 Masking history 2 Advantages: •Quicker recovery •Less cost clean T[i1] T[k1] T[j1] •If T[i1], T[j1], and T[k1] are all malicious, the main database is valid •If T[i1] and T[j1] are malicious, but T[k1] is not, then masking engine 2 is valid •If T[i1] and T[k1] are malicious, but T[j1] is not, then though none is valid, reexecuting T[j1] on the main history can produce the valid database 26 A Limitation of Scheme 4 •Scheme 4 is not adaptive by nature •Adaptation can give better resilience and cost-effectiveness •There is no automatic way for the system to adaptively adjust its defense behavior according to: •the characteristics of recent and ongoing attacks •its current performance against these attacks •Although the SSO can dynamically reconfigure some of its components, manual reconfiguration operations are ad-hoc, not scalable, and prone to errors •Our goal: adaptive database intrusion tolerance 27 Scheme 5: Self-Stabilization •Self-Stabilization: the degree of data integrity should be able to be automatically stabilized within a tolerable range no matter how the system is attacked User SQL Commands Damage Confinement Mediator (Policy Enforcement) Intrusion detector Damage Assessor Damage Repairer Tolerable range State variable feedback The controller Main database Isolation engines Masking engines 28 The database Optimized Load Balance •Observation: •Different load configurations usually cause different cost-effectiveness •A load configuration can cause very different cost-effectiveness in different situations •An example of load configuration: •the percentage of isolated users •the percentage of masked users •the percentage of malicious users •the number of masking engines used •the average interval of state variable feedback •... •Our goal: adaptive load configuration optimization •Mechanism: the controller can be responsible for this job 29 New Progress of Scheme 5 • Since Feb 2001 • We have investigated rule-based (adaptive) selfstabilization techniques • Some example self-tuning rules are produced • Overall 30 Metrics to measure success •Cost –time, space needed for tolerating intrusions •Effectiveness –how many intrusions are detected, isolated, or masked –how many mistakes are made –how effectively can the damage be confined –how quick can the damage be assessed and repaired –how well can the system be adapted –availability: how often is a legitimate request rejected –integrity: how well can data integrity be preserved under attacks •Performance –system throughput –response time 31 Task Schedule Repair Mediator Detection Confine Isolation Masking Self-Tuning Design Implementation 100 90 80 70 60 50 40 30 20 10 0 Evaluation 32 Technology Transfer •Technical papers published in leading technical meetings and technical reports • Release and dissemination of the prototype in source and binary forms •Pursuing technology transition through major commercial DBMS vendors. The technologies can either be absorbed into their DBMS kernels, or be commercialized as intrusion tolerance wrappers •Starting a company to commercialize the technologies and provide flexible services to arm customers' database systems with necessary intrusion tolerance facilities 33 Questions? Thank you! 34 Multi-layer representation of our approach 35