Semantic Modeling & Monitoring for Real Time Decision Making: Results and Next Steps within the Greek Cyber Crime Centre of Excellence (GCC)
- 1. ΚΕΝΤΡΟ ΜΕΛΕΤΩΝ ΑΣΦΑΛΕΙΑΣ CENTER FOR SECURITY STUDIES UKSIM-AMSS: 15th International Conference on Modelling and Simulation Cambridge University (Emmanuel College), 10-12 April, UK. Semantic Modeling & Monitoring for Real Time Decision Making: Results and Next Steps within the Greek Cyber Crime Centre of Excellence (GCC)
- 2. Team Presenter: Vasilis Tsoulkas, PhD, Systems Analysis. Research Group: Dimitris Kostopoulos, MSc, Software Analyst George Leventakis, PhD, Security Risk Analyst Prokopis Drogkaris, PhD, Security Policy Analyst Vicky Politopoulou, MSc, Forensics Analyst Parts of this research implemented in cooperation with : the IT Innovation Center, Southampton , UK. 2
- 3. Presentation Sections ● Motivation & Objectives – FP 7 funded - SERSCIS Proof of Concept Architecture - successfully delivered to the EC ● Semantic System Modeling Aspects (The Dynamic Multi-Stakeholder Approach) ● Semantic Monitoring and Stream Reasoning Process ● Decision Support Tool Interfaces & Risk Analytics ● Next Steps within the Greek Cyber Crime Center of Excellence (GCC). 3
- 4. Motivation and Objectives ● The FP7 Project addressed the problem of Dynamic Modeling and Dynamic Risk Management in Critical Infrastructures (Cis). ● Today’s CIs are characterized by: Increased Connectivity (between different Stakeholders) of their Information and Data Processing Infrastructures ● ….for Increased Performance and Efficiency ● …but this situation introduces new challenges of Cyber-Physical Threats and their Counter-Measures 4
- 5. Motivation and Objectives 3 main causes for increased CI vulnerabilities 1.Cyber-Attacks against interconnected Information & Communication channels can disrupt Exchanged Data flow and Integrity 2.Local Disruption in one Organization (stake holder) generates problems elsewhere due to coupling and dependencies (of data and networks). 3.Reduced Resilience against any cyber-disruptions due to reduced excess capacity arising from the exchanged data for increased efficiency requirements 5
- 6. Objectives ● Implementation of Agile Service Oriented Technologies to ● compose ICT connections related to the critical infrastructure ● monitor and manage ICT components against well-defined dependability criteria ● adapt ICT connections in response to disruption or threats ● validation of this approach in Proof of Concept Scenarios from the air traffic sector (Airport-Collaborative Decision Making (A-CDM) /EUROCONTROL) 6
- 7. EUROCONTROL / Airport-Collaborative Decision Making – European Air Traffic Management System Data Exchange Service accessible by a consumer (aircraft operator) through SLA template consumer. The GH is responsible for coordination of Ramp Services (catering, fuelling, cleaning, baggage handling) GH: Is an Orchestrator of Ramp Services to have an aircraft ready for its next flight 7
- 9. Data quality and KPIs in A- CDM ● Data: Confidentiality, Integrity, Alarms, Data Display ● KPIs: Reflect the Quality of Service Delivery ● KPIs data properties: Is the Quality of Time Estimates Accuracy Predictability Stability ● An SLA Architecture was developed with the following KPIs & Parameters in the Airport Collaborative Decision Making (A-CDM) context: • System Availability • Data Quality • Data timeliness, delivery deadlines • Confidentiality 9
- 10. Performance of Service Providers & KPIs ● The GH performance is characterized by two KPIs: TOBT (Target Off Block Time) i) accuracy and ii) stability; i) TOBT accuracy parameter: It is derived by comparison with a Ref. Value : Actual Ready Time (ARDT). The Mean SQ. Deviation between TOBT & ARDT is computed for all flights departing in a day. ii) TOBT stability parameter : Measures how stable is the predictor mechanism of the GH (estimates). • The two KPIs affect the number of take-offs outside a Slot Tolerance Window (STW). • Another KPI for overall CDM performance evaluation: Average Number of Slots / flight. (Ideally: One Slot/Flight) 10
- 11. Addressing the Modeling Challenge for Machine Reasoning For the Dynamic Multi-Stakeholder system we constructed : 4-levels of abstraction 1. A core (ontology) structure: to model the System and its assets subject to threats and protected by controls 2. A dependability model: describing system independent: assets, threats, controls using OWL classes and relationships. Security expertise is encoded in this model. 3. An abstract system model: describes system-specific threats and controls. Extends the dependability model classes with security knowledge. 4. A concrete system model: provides a snapshot of the running system and instances of the participating assets + contextualised threats & controls. Each level inherits from its predecessor (parent – child relation). The final concrete model has simple structure and integrates knowledge from: Abstract system model and Dependability model. 11
- 12. Brief Analysis of Ontology & Models 1. The Semantic Ontology is constructed such that: ● Only OWL Classes are used for design-time modelling ● OWL Instances are used for modelling the Run – Time System Composition ● Security expertise is added at design time in the OWL classes 2. The Dependability model provides the first step to develop the Abstract System Model which is a Design – Time Model of the system that will be composed dynamically “On the Fly” (run- time). 3. The Concrete Model Generator is connected to the monitoring subsystem to create a model of the Running System (Current System Composition). 12
- 13. Main Innovation of the Approach ● The Modelling approach is constructed using Semantics Modelling for Machine Reasoning automated threat analysis and risk estimation when the system is composed at Run-Time. ● The design – time Service Oriented Dynamic models are abstract: They describe the structure but NOT the composition of the system which is NOT KNOWN until “Run-Time”. 13
- 14. Core System Domain Ontology ● This basic system structure, determines what reasoning is used Threat class Description Controls needed Unauthorized The service processes an unauthorised request Client AuthN + Client access from an attacker. AuthZ Unaccountable Type of unauthorized access, designed to get Client AuthN + Client access the service without paying for it. AuthZ 01/02/2013 14
- 15. Dependability System Model (High Level View) Threat impact is Threat activity may affect described in terms of the behaviour of the the intended (or asset, or controls on the unintended) asset may reduce the compromise threat threatens Threats are sub- classed to create Assets are sub-classed 1 generic and system- to create generic and specific threat types system-specific asset affects types Asset Threat * 1 * blocks/ protects mitigates Control rules define depends which controls block on or mitigate the Control compromise of the threatened asset Control presence is Controls are sub- determined by asset classed to create behaviour generic control types only 15
- 16. Dependability System Model (Controls & Threats) Dependability Model Logical Asset Types & Relationships A generic model of dynamic, multi-stakeholder service-oriented systems including generic threats and threat classification (control) rules 16
- 17. Threat Types & Threat Scenario 1. Unauthorized Access (to the service) 2. Data traffic Snooping 3. Man in the Middle 4. Client Impersonation 5. Resource Failure Unauthorized Data Update at Fuelling Service 17
- 18. Dependability Model Controls (sample ) ● Controls provide defences against generic Threats Two broad categories : ● proactive controls block a threat, against the protected asset ; ● reactive controls allow the effect of a threat on the asset to be mitigated once the threat is carried out . ● Mitigation and Blocking rules are of the form : ThreatClass(?t) ∧ AssetClass(?a) ∧ threatens(?t,?a) ∧ ControlClass1(?c1) ∧ protects(?c1, ?a) ∧ … ∧ ControlClassN(?cN) ∧ protects(?cN, ?a) → MitigatedTreat(?t) ThreatClass(?t) ∧ AssetClass(?a) ∧ threatens(?t,?a) ∧ ControlClass1(?c1) ∧ protects(?c1, ?a) ∧ … ∧ ControlClassN(?cN) ∧ protects(?cN, ?a) → BlockedTreat(?t) 18
- 19. Threat Block – Mitigation Rule Explanation ThreatClass(?t) ∧ AssetClass(?a) ∧ threatens(?t,?a) ∧ ControlClass1(?c1) ∧ protects(?c1, ?a) ∧ … ∧ ControlClassN(?cN) ∧ protects(?cN, ?a) → MitigatedTreat(?t) or ThreatClass(?t) ∧ AssetClass(?a) ∧ threatens(?t,?a) ∧ ControlClass1(?c1) ∧ protects(?c1, ?a) ∧ … ∧ ControlClassN(?cN) ∧ protects(?cN, ?a) → BlockedTreat(?t) ● First line in each case finds a threat instance of a specific type that threatens an asset instance of a specific type. ● Second line looks for control instances of the right types to protect the threatened asset against this type of threat. ● Last line classifies the threat as either mitigated or blocked, depending on the types of controls affording this protection 19
- 20. Control Class Explanation Control classes provide: ● generic control types that can be included directly in an abstract system model; ● descriptions of deployment actions: how to deploy the control into the real system; ● descriptions of mitigation actions: how to operate reactive controls to protect assets when a threat is carried out against them. 20
- 21. Resource Software Malfunction (Mild Error) ● Threat Provider- Specified ● a bug in PSResource software Resource causes it to repeatedly produce faults threatens protects ● Controls ● PSResource has Suplier blocks Software Patching: the Supplier Resource Software Software Supplier has a procedure to Malfunction Patching maintain the software used by the PSResource ● ensures bug fixes are applied promptly ● System specifics ● one subclass per PSResource class ● one instance per PSResource of each of the resulting classes
- 22. Abstract System Model of A- CDM ● It is a design-time model of the structure of the dynamic, multi-stakeholder service-oriented system: Input for fully automated run-time model generation and analysis Tools. It is composed dynamically at Run-Time. 22
- 23. Concrete System Model in the Overall Architecture ● A run-time model of the system, including its current composition, status and threat-related behaviours (Current System Composition) ● It is Created automatically, and it is used to provide run- time decision support 23
- 24. Classification / Estimation Blocks - Output Concrete Model : Input for a) Threat Classification & b) Threat Likelihood Estimation The output of these two tools provides: • A list of potential threats with classification • Estimates of the likelihood a threat is active or in-active • A description of each Threat with impact severity • A list of controls to block or mitigate a threat and if these controls are available in the system 24
- 25. Semantic Monitoring & Stream Reasoning Process (Initial Prototype & Limitations) This initial monitoring – reasoning process has 3-stages: Starting point: Semantic model of the structure of the dynamical multi – stakeholder system. The model was “Abstract” based only on OWL Classes with “types of entities” but NOT YET “Entity Instances”. These are only known at Run-Time. !!! At Run-Time the Concrete System Model was Generated with OWL Instances reflecting the System Composition. Concrete Model generation used Status Monitoring Data from Run Time Monitoring and Management Block 25
- 26. Some…..Limitations of this Approach ● Run – Time Concrete model reflects instantaneous snap-shot of System Composition, Behavior & Status. No past Information. !! ● Threat Analysis depends on Current Concrete Model and NOT on past Snap-Shots (History). !! ● Threat activity is computed from current concrete system model NOT on past Snap-Shots (History). !! ● No Correlation between Activity and different threats 26
- 27. Some…… PoC Implementation Limitations ● A complete set of monitoring data is needed to generate the Concrete System Model Snap-shot with Re-Generation of even static features of the model each time. !!! The A-CDM PoC. ● This generation of the Complete - Concrete Model is “Time Consuming”. Every new model is “out-dated” !! missing intermediate “rapid changes information”. • Non-Distinguishability between asset – compromises (Behaviors) that produce the same instantaneous monitoring data even for different Past Monitoring data. 27
- 28. New Approach: Stream Reasoning ● • Information arrives as a stream of “time-stamped” data • The Knowledge base can be continuously updated and reasoning goals are continuously re-evaluated as new assertions arrive • Reasoning is implemented from a Finite – Time Window and not at a Single Instant !!. • Research Efforts on Stream Reasoning is still at its First Steps and at its Infancy. 28
- 29. 3 basic – steps in Stream Reasoning ● Select: Relevant Data from Input Streams by using Sampling Policies that probabilistically drop stream elements to address bursty streams of data (unpredictable peaks). ● Abstract: Sampled streams are input to Abstract block to generate aggregate events by enforcing aggregate events continuously. Output is RDF streams (ρ, τ) with ρ – RDF triple and τ – time stamp (logical arrival time of RDF statement. Use of C-SPARQL. ● Reason: RDF (Graph Streams) streams are injected into background knowledge for reasoning tasks. Incremental implementation of RDF snapshots. 29
- 30. Semantic Monitoring Block Semantic Reasoning Block ….excluded 30
- 31. Semantic Monitoring Component : DSMS - Behavior Analyser - Sequential Detection DSMS: Data Stream Management System : samples & filters monitoring data generated by Service Monitoring and Management Components. ● Usage of open-source CEP (Java - ESPER): Real Time engine that triggers Listeners or Subscribers using a tailored Event Processing Language (EPL). 31
- 32. Behavior Analyser (BA) ● Processing of multiple data streams from DSMS. Produced Output is Graph Triples (RDF). • Decides how to convert raw monitoring data into Semantic Assertions related to: Presence of Assets and Behaviors. ● The monitoring framework generates 2 – types of Time stamped RDF assertions: (1) Presence or Absence of Assets (joining or leaving the system) (2) Assertions about Measurability, Presence or Absence of Adverse Behavior of these Assets. 32
- 33. Behavior Analyser (BA) ● The BA is not only a Transcoder converting Monitoring Events to RDF graphs. ● The BA decides about the type of Behaviors (Assets and Services). ● Example: The BA is capable to determine if an Asset is Overloaded or Underperforming using Monitoring Data for Load and Performance (KPIs – SLA events). 33
- 34. Sequential (Behavior) Detection Well – Known Cumulative Sum (CUSUM) algorithm from the sequential statistics literature. In general parametric models are used Change in the mean of the relevant stochastic process We use: The non-parametric version of CUSUM 34
- 35. Sequential (Behavior) Detection Mean Value Attack Initiation Time and Detection Delay Test Statistic 35
- 36. Non-parametric CUSUM: Mathematical principles Random Sequence yn =( yn −1 +Z n ) + Test Statistic y0 =0 X + = max(0, x ) If We Define: The max Continuous Increment up to time n. 0 : Normal Operation Decision Making Rule: 1 : Attack Event: N is Threshold 36
- 37. DST – Tool Dynamic Interfaces Scenario: Remote exploitation on Fuelling Services GH Fuelling Service PSR_InterruptedCommunication Host Group X scheduling Airport NET Remote exploit Attacker AttacK: Attacker on the AirportNet network targets the Host of the Fuelling Service. RKE: Remote Known Exploit 37
- 38. SERSCIS DST interface (Logical Assets View) 38
- 39. DST interface (Physical Assets View) 39
- 40. DST interface - Assets Under Threat 40
- 41. DST interface (Threats Involving Selected Asset) 41
- 42. DST interface (Threat Information and Countermeasures Proposition - Zoom) 42
- 43. EU funded Project GCC: A Cyber Crime Center of Excellence for Training, Research and Education in Greece 43
- 44. Objectives ● To create a Greek Cyber Crime Centre of excellence & develop a sustainable infrastructure for GCC. ● To mobilize the Greek constituency in the area of cyber crime training, research, and education. ● To advance research in the area of cybercrime, focusing particularly in areas dealing with ● cyber attacks, ● botnet research ● Intrusion detection systems ● Intrusion detection systems for Critical Infrastructures. 44
- 45. Work Packages GCC Project (36 months) Starting 1st January 2013 WP1: Management WP2: Dissemination WP3: Education and Training WP4: Research 45
- 46. Work Packages GCC Project FORTH AUTH SAFENET KEMEA (WP3) (WP1) (WP2) Training (LEAs, Management Dissemination Website(pub./priv.) (WP3) Private/Public Advisory Board University 2CENTRE - CCUs sector) (WP3) courses Repository (WP4) (University, LEAs) Research (WP4) (WP4) Research (Intrusion (WP4) Research Legal issues (legal Detection System (Disruptive framework, –Critical monitoring, Botnet Policy) infrastructure Detection tool) taxonomies) 46
- 47. GCC ● A CyberCrime Center of Excellence for Training, Research and Education in Greece” . ● Funded under Prevention of and Fight against Crime (ISEC) Programme of European Commission Directorate-General for Home Affairs. ● GCC main objectives are to create a constituency of Greek stakeholders in the area of Cyber Crime and to mobilize this constituency to work together in education and research areas. 47
- 48. Next Steps in the GCC framework ● Exploitation of sequential detection of a change using the nonparametric CUSUM in the Behavioral Analyzer. ● Situational Awareness of the Operators using user friendly Dynamic Support Tool (DST) interfaces ● Further development using additional detection approaches (Sequential Probability Ratio Test, Different Optimality Criteria such as: Lorden, Shiryaev - Roberts) ● Distributed Real Time Sequential Detection & Hypothesis Testing for Intrusion Attacks ● Some Application areas: Industrial CIs Protection, Network Intrusion Detection Systems, Swarms & Networks of cooperative UAVs. 48
- 49. Conclusions ● Implementation of an Intelligent Prototype Tool for the Protection of Dynamic Multi Stakeholder SOA Critical Infrastructures. Air-traffic Management Systems PoC. ● Implemented: An Innovative core ontology model which has been reinforced with rules and classes that improve threat estimation and classification. ● Implemented: Advanced Stream (RDF) Reasoning – and Behavioral Analysis Algorithms. ● Sequential data analysis led us to Advanced Semantic Stream Reasoning for Real –Time Processing. ● Implemented: Dynamic User Interfaces with Risk – Threat Analytics in Real Time for A-CDM (Eurocontrol). 49
- 50. Questions – Discussion. Thank you ! Contact Details: Vasilis Tsoulkas tsoulkas.kemea@gmail.com Dimitris Kostopoulos dimkostopoulos@gmail.com George Leventakis george.leventakis@gmail.com Prokopis Drogkaris prokopis.drogkaris@gmail.com Viky Politopoulou v.politopoulou@gmail.com 50
Editor's Notes
- 01/02/2013 George Leventakis, KEMEA