SRS Common Architecture Bob Balzer Neil Goldman Dave Wile Teknowledge Corp.

18
SRS Common Architecture Bob Balzer Neil Goldman Dave Wile Teknowledge Corp.

Transcript of SRS Common Architecture Bob Balzer Neil Goldman Dave Wile Teknowledge Corp.

Page 1: SRS Common Architecture Bob Balzer Neil Goldman Dave Wile Teknowledge Corp.

SRS Common Architecture

Bob BalzerNeil Goldman

Dave WileTeknowledge Corp.

Page 2: SRS Common Architecture Bob Balzer Neil Goldman Dave Wile Teknowledge Corp.

• Defense In Depth– Outer Prevention Layer: Block most attacks

Prevention

SRS Integration Architecture

Proactive Prevention

Reactive Prevention

– Inner Regenerative Layer: Repair (and prevent) unblocked attacks

Self-RegenerationRecovery

Health MonitoringDiagnosis

Memory Layout Diversity, Network Filter, Scalable Redundant Storage, Robust Scalable Comm.

Software Dynamic Translation, Generated Network Filter, Dynamic Method Dispatch

Component Diagnosis, Attack Recognition, Malicious Intent Determiner

Corruption Extent Determiner, Reconstitution Planner

Architecture Diffferencer, Harm Detector

Senso

r

Repair

Page 3: SRS Common Architecture Bob Balzer Neil Goldman Dave Wile Teknowledge Corp.

Technical Goals

• Include as many SRS program elements as possible

• Minimal intrusion into existing tools, primarily – by announcing system status incrementally

– And dynamically responding to other tools’ results

• Integrate capabilities for seamless interoperation• Stimulate the production of new capabilities to

further integration goals

Page 4: SRS Common Architecture Bob Balzer Neil Goldman Dave Wile Teknowledge Corp.

Project Abbreviations

• AWDRAT: MIT: Shrobe and Teknowledge:Balzer

• Cortex: Honeywell: Musliner

• Daikon: MIT: Ernst and Rinard

• Dawson: Global Infotek: Just

• JHU: Johns Hopkins U.: Amir and Purdue U.: Nita-Rotaru

• MBE (model-based executive): MIT: Williams and Sullivan

• PMOP: Teknowledge: Balzer and MIT: Shrobe

• SensorNet: Telcordia: Van Den Berg and Rutgers: Rajagopalan

• Strata (Genesis): UVA: Knight, Davidson, Evans, Nguyen-Tuong and CMU: Wang

Page 5: SRS Common Architecture Bob Balzer Neil Goldman Dave Wile Teknowledge Corp.

SRSComponent

Shared System Architecture

Announce and analyze system status

DawsonPMOP

JHU

DaikonSensorNet

Strata

MBE

Cortex

AWDRAT

Chooseresponse

Effectresponse

Page 6: SRS Common Architecture Bob Balzer Neil Goldman Dave Wile Teknowledge Corp.

Technical Approach

• Parallel monitoring and analysis by SRS components of a single target system

• Components communicate via a global blackboard

• Blackboard organized by a shared ontology for describing system and heartbeat states

• Subscriptions provide access to others’ sensors, analysis, and response choice

Page 8: SRS Common Architecture Bob Balzer Neil Goldman Dave Wile Teknowledge Corp.

Overview of Potential ScenariosSetup

Install Learning Harness (Cortex, Daikon, MBE, PMOP)

Determine Method / Class Equivalents (PMOP, Strata)

Install Wrappers and Obfuscate DLLs (AWDRAT, Dawson, PMOP)

Data Error (Daikon)

Program Error (AWDRAT, Cortex, Dawson, JHU, MBE,

Strata)

Operator Induced Error (PMOP,

SensorNet)

DetectHeartbeat

(Cortex)

Attack Indicator (Dawson)

Collateral Damage Assessment (AWDRAT,

Daikon, JHU, MBE)

Trust and Risk (AWDRAT, Cortex,

MBE, PMOP, SensorNet)

AnalyzeAssess Learning Data

(Cortex, Daikon, MBE, PMOP)

Data Error:

Data Repair (Daikon)

Restore Data / DB (Cortex)

Operator Induced Error:

Automated surrogate or backup operator

Omit the affected component

Propose RepairProgram Error:

Execute a component in the virtual machine (Strata)

Select from method alternatives & regenerate from scratch

(AWDRAT, Dawson, MBE)

Select from method alternatives & backtrack (JHU)

Page 9: SRS Common Architecture Bob Balzer Neil Goldman Dave Wile Teknowledge Corp.

Scenarios Continued

Choose Repair(New Work to choose among additional proposed repairs)

Data Error:

Data Repair (Daikon)

Restore Compromised DB

(Cortex)

Program Error:

Execute a component in the virtual machine (Strata)

Select from method alternatives & regenerate from scratch

(AWDRAT, Dawson, MBE)

Select from method alternatives & backtrack (JHU)

Make RepairOperator Induced

Error:

Prevent bad effects (PMOP, SensorNet)

Ignore or encapsulate the error (New Work)

Page 10: SRS Common Architecture Bob Balzer Neil Goldman Dave Wile Teknowledge Corp.

Blackboard Organization

Blackboard layers correspond to scenario layers• S (setup)

• D (detect)

• A (analyze)

• PR (propose repair)

• CR (choose repair)

• MR (make repair

Page 11: SRS Common Architecture Bob Balzer Neil Goldman Dave Wile Teknowledge Corp.

Messages Passed• Setup and Status SRS Agents

– S:Environment attribute or input A has value V– S:Program mode for Sys is M– S:System components for Sys are { ci } – S:Variants for CID are { ci }

• S:Variant generator for CID is SRSAgent [ mode, CID]– S:Checkpoint Sys in D– S:GUI checkpoint E in D

• Detection SRS Agents– D:Program Sys had fault F at L in ci [where L is contained in Sys]

(Fault is supertype of DataError, OperatorInducedError, ProgramError)– D:Missed heartbeat Sys at time T fault F– D:Attack of Sys indicator I for CID at L in ci

• Analysis SRS Agents– A:Program Sys has vulnerability V at L { to risk R }– A:Program Sys has collateral damage V at L– A:Component ci would incur risk R with certainty P– A:End of Positive {or Negative} learning example trial for Sys

Page 12: SRS Common Architecture Bob Balzer Neil Goldman Dave Wile Teknowledge Corp.

• Propose Repair, Choose Repair & Make Repair SRS Agents – Same messages used in each layer for different purposes:– Detectors and analyzers populate PR layer with these assertions (see

following Messages).

– Choose Repair agent asserts these same facts into CR, triggering repair.

– Effectors assert these facts into MR to indicate repair completion.

• Messages– [layer] :Replay component ci from checkpoint D in history H

– [layer] :Substitute ci in Sys at L [used for Data repair, database substitution, and program regeneration]

– [layer] :Remove component ci in Sys at L

– [layer] :Revert Sys to checkpoint D in history H

Messages Passed

Page 13: SRS Common Architecture Bob Balzer Neil Goldman Dave Wile Teknowledge Corp.

Ontology for Blackboard

• All blackboard objects part of ontologically described database

• Historical and Metadata – facts as objects

– May need special assertions

– And special queries

Page 14: SRS Common Architecture Bob Balzer Neil Goldman Dave Wile Teknowledge Corp.

OntologyDemo

Page 15: SRS Common Architecture Bob Balzer Neil Goldman Dave Wile Teknowledge Corp.
Page 16: SRS Common Architecture Bob Balzer Neil Goldman Dave Wile Teknowledge Corp.

Blackboard Design Issues• System representation

– Identity– Versions (especially with learning)– Historical data– Specific types

• Programs• GUI actions• States• Environment • Control Modes

• Conflict resolution• Control resolution

– Metadata– Layered blackboard– Agent relationships

• Activities and Results to be communicated among SRS agents

Page 17: SRS Common Architecture Bob Balzer Neil Goldman Dave Wile Teknowledge Corp.

Traditional Conflict Resolution Solutions

• “First” rule by some criterion• Highest “priority” rule• Most “specific” rule• Rule that refers to the element most “recently”

added• New rule• Arbitrary• All rules in parallel• Compartmentalized knowledge

Page 18: SRS Common Architecture Bob Balzer Neil Goldman Dave Wile Teknowledge Corp.

Race Conditions• Blackboards lose their elegance when agents cannot freely access them, e.g.

when agents don’t know whether to wait for more information to arrive.– E.g. Good: agents A and B both analyze message M and report their results

independently– E.g. Bad: agents A, B, and C all analyze message M but B and C need A to have

“passed” message M before they can work– Bad solution: B waits for A to “bless” M or “fail M” before proceeding

• Good solution: when A can respond to M in a blackboard layer not examined by B and C, subsequently asserting M into a blackboard layer that both B and C look at.

• If they’re all in the same layer, a possible solution is lattice-based access within a layer:

– Register B and C as higher in the layer’s lattice relating all agents’ access– When a message M arrives that A is interested in, it is sent to A first.– If A reasserts M, both B and C can act on it. Otherwise, it has been “consumed”

and must be removed from the blackboard.– Multiple, simultaneous messages: prefer complex message groups to simpler ones.

• Use to introduce fault analyzers and repairchoice between layers. B C

A