Composed Statistical Modeling and Optimization for Ensuring ... .


28 views
Uploaded on:
Category: General / Misc
Description
Coordinated Statistical Modeling and Optimization for Ensuring Data Integrity and Attack-Resiliency in Networked-Embedded Systems. Farinaz Koushanfar, ECE Dept. Rice University Statistics Colloquium Oct 9, 2006. outline. Sensor Networks: Applications, Challenges
Transcripts
Slide 1

Facilitated Statistical Modeling and Optimization for Ensuring Data Integrity and Attack-Resiliency in Networked-Embedded Systems Farinaz Koushanfar, ECE Dept. Rice University Statistics Colloquium Oct 9, 2006

Slide 2

plot Sensor Networks: Applications, Challenges Coordinated Modeling-Optimization Framework Inter-sensor models Embedded detecting models Optimization for information respectability Attack Resilient Location Discovery Problem plan and assault models Robust irregular specimen agreement for assault identification Sensor Networks: Applications, Challenges

Slide 3

Air contamination Texas wine!! http://www.ucsusa.org/clean_energy/coalvswind/c02c.html Vibration in Abercrombie http://www.alamosawinecellars.com/vineyard2.htm Flood in Houston http://dacnet.rice.edu/maps/space/index.cfm?building=abc http://www.bluishorange.com/surge/photographs/10bridgecars.jpg Sensor Networks Comprehensive observing and examination of complex physical situations Imagine…

Slide 4

Sensor Networks, How? Systems of installed detecting (activating) and registering gadgets Mica2Dot, CrossBow Tech. Cordiality of Prof. Estrin, CENS, ULCA

Slide 5

Challenges in Sensor Networks System: sensors, actuators, equipment, programming, correspondence organize layers, Limited: battery, transfer speed, cost Unique to sensor systems: Sensing Abstract the framework state, complex properties, and model physical marvels precisely, without inclinations Parametric models: from the earlier suspicions Often don\'t catch the mind boggling connections Optimization in light of such models have a restricted viability

Slide 6

Challenges in Sensing Massive datasets Structure reaction in USGS building: 72 channels of 24 bit information, 500 examples/sec. Vitality utilization of the remote hubs Motes take 36mW in dynamic mode ��  AA batteries + capacity limit of 1850mWh  50h dynamic mode Diversity in applications Marine science, seismic detecting, combat zone, contaminant transport, home sensors, labs, healing centers, and so forth. Unforgiving natural conditions Battlefield, seismic tremors, programmed location, and so on. Remote channel information misfortune Sensor cost Sensitivity of utilizations Privacy and security

Slide 7

Inconsistencies in the Measured Sensor Data Erroneous estimations Noisy readings: inescapable because of influence and cost requirements and natural effect Systematic blunders: counterbalance predisposition, adjustment impact, and so forth Partially tainted, still valuable Faulty (adulterated) estimations Remove deficiencies to get a reliable picture Can be inadvertent (e.g. terrible connection), or malignant Missing information May be unintentional, purposeful (resting, subsampling, pressure, sifting), or vindictive

Slide 8

Outline Sensor Networks: Applications, Challenges Coordinated Modeling-Optimization Framework Inter-sensor models Embedded detecting models Optimization for information trustworthiness Attack Resilient Location Discovery Problem plan and assault models Robust irregular specimen accord for assault recognition Coordinated Modeling-Optimization Framework

Slide 9

Motivational Example Deployments demonstrate a hole b/w models and the truth Example: preparatory investigation of temperature sensor follows at UCLA BG 23 sensor hubs, testing every 5 mins Question: does the area presumption hold?

Slide 10

Motivational Example (Cont\'d) No predictable connection b/w detecting and separation Discontinuities, presentation contrasts, worldwide sources Also, some profoundly corresponded near to sensors Best past exertion: neighborhood premise capacities Need new models for concurrent reflection of detecting and separation What about different properties?

Slide 11

Motivational Example (Cont\'d) Separation of concerns Embedded detecting models: Define various charts G 1 , G 2 , … , G M , that share vertices E.g., detecting, remove d ij : remove b/w s i ,s j e ij : detecting forecast blunder, for the model s j =f ij (s i ) The separation and detecting are not stuck into one model, but rather are as a rule at the same time considered

Slide 12

Motivational Example-2 Cross-space streamlining: Sensor sending Objective: select up to S hopeful focuses for including an additional sensor For every s i , a TL sensor is Delaunay neighbor yet can\'t be anticipated inside  th mistake bound Denote the edges of TL sensors as applicants Find smart approaches to choose the best arrangement of competitor focuses

Slide 13

Motivational Example (Cont\'d) Coordinated demonstrating advancement Q1: How to do cross-area enhancement? Q2: Can the models be of higher measurements? Q3: Can they help us to address information uprightness issue? Q4: How viable would they say they are?

Slide 14

Outline Sensor Networks, Applications, Challenges Coordinated Modeling-Optimization Framework Inter-sensor models Embedded detecting models Optimization for information respectability Attack Resilient Location Discovery Problem plan and assault models Robust irregular specimen accord for assault identification Inter-sensor models

Slide 15

Inter-sensor Models Intra-sensor models (autoregressive models) Have demonstrated the adequacy of adding shape requirements to univariate models Isotonicity Unimodularity Number of level sets Convexity Bijection Transitivity Combinatorial isotonic relapse (CIR), finds the ideal nonparametric shape compelled univariate fit for a self-assertive mistake standard in normal straight time Models are forerunner for consequent improvement

Slide 16

Application of CIR on Temperature Sensors at Intel Berkeley * Prediction blunder over all hub sets Limiting the quantity of level sets * Koushanfar, Taft (Intel), Potkonjak Infocom\'06

Slide 17

Multivariate CIR Recent result * : The principal ideal, polynomial-time DP-based approach for multi-dimensional CIR: (1) Build the relative significance lattice R (2) Build the blunder grid E (3) Build an aggregate mistake grid C by utilizing a settled DP (4) Starting at the base esteem in the last section of C, follow back a way to the primary segment that minimizes the total blunder Thanks to Prof. D. Brillinger (UCB), Prof. M. Potkonjak (UCLA) for the helpful dialogs

Slide 18

Example

Slide 19

Multivariate CIR - many-sided quality T sensor values drawn from a limited letters in order A Complexity of univariate case is ruled by sorting (T log T) C (M): multifaceted nature of multivariate with M logical factors C m (M)=A M+1 C (M-1), pseudo-polynomial unpredictability

Slide 20

Open Questions How to accelerate the Multivariate CIR? Pruning calculations that adventure sparsity (?) Is it conceivable to make CIR locally versatile ? On a fundamental level, finding the min blunder is a worldwide advancement that can\'t be privately tended to Can one ensure union and accuracy of CIR among sensors? Is it conceivable to have ceaseless approximations to address the issue? In what capacity would one be able to assemble effective models in nearness of missing as well as defective information?

Slide 21

Outline Sensor Networks, Applications, Challenges Coordinated Modeling-Optimization Framework Inter-sensor models Embedded detecting models Optimization for information uprightness Attack Resilient Location Discovery Problem detailing and assault models Robust irregular specimen agreement for assault recognition Evaluation and correlation with contending techniques Embedded detecting models

Slide 22

State-of-the-Art Sensing Models Parametric models Gaussian arbitrary fields, graphical models (GM), message passing, iterative message passing, conviction proliferation (BP) Nonparametric models Marginalized portions (GM), rotating projections, circulated EM, nonparametric BP Common string: catch reliance among sensor information, no edge implies no reliance, Need to catch the state of field discontinuities and additionally absence of relationships b/w nearby hubs

Slide 23

Embedded Sensing Models Principle of partition of concerns (SoC) Example: Geometric diagram (planar-2D) Delaunay edges (nearness) Sensing chart: higher dimensional installed chart 1 5 4 7 2 6 2 8 1 6 3 8 5 7 4 Idea: Map the detecting diagram into lower measurements. Abuse the error between the higher dimensional topology and the lower dimensional space to distinguish the snags

Slide 24

Open Questions Efficient calculation and treatment of implanted detecting models in higher measurements Joint pressure of various elements How would we be able to catch dynamic topologies , i.e. portability, dynamic time arrangement, dozing Efficient structures/information designs for speaking to the multi-dimensional topologies

Slide 25

Coordinated Modeling and Optimization Paramount significance of interface in framework and programming advancement Create factual models reasonable for improvement Paradigms: nonstop, smooth, reliable Small number of level sets Convexity Bijection x\' i = G(F(x i )) = x i , where y i =F(x i ) and x i =G(y i ) Transitivity z i = F(x i ) = G(y i ) Create streamlining components flexible to measurable inconstancy Paradigms: randomization Multiple approvals Constructive probabilistic Reweighting OF and requirements

Slide 26

Outline Sensor Networks, Applications, Challenges Coordinated Modeling-Optimization Framework Inter-sensor models Embedded detecting models Optimization for information uprightness Attack Resilient Location Discovery Problem plan and assault models Robust arbitrary example agreement for assault recognition Evaluation and correlation with contending techniques Optimization for information honesty

Slide 27

Data Integrity: Multiple Validations The information respectability issues are unpredictable because of the mind boggling situations and vulnerabilities Proof of NP-culmination (PhD\'05) Data trustworthiness (commotion lessening, alignment, blame identification, information recuperation) misuses framework redundancies Coordinated demonstrating enhancement Multiple approvals (MV) advancement calculations The arrangements are approved utilizing various info tests Similar as a part of soul to cross-approval (CV) in measurements MV is more exhaustive than CV, since it is a bland improvement worldview in light of resampling the information space and approving the yield of a perplexing calculation rath

Recommended
View more...