Transcript Bad Data Injection in Smart Grid - Wireless networking, Signal

Bad Data Injection in Smart Grid:
Attack and Defense Mechanisms
Zhu Han
University of Houston
Overview

Introduction to Smart Grid

Power System State Estimation Model

Bad Data Injection

Defender Mechanism
– Quickest Detection

Attacker Learning Scheme
– Independent Component Analysis

Future Work

A Few Topics in Smart Grid Communication

Conclusions

Quick View of Amigo Lab
“Smarter” Power Grid

Sensing, measurement, and control devices with two-way
communications between the suppliers and customers.

Benefits both utilities, consumers & environment:
– Reduce supply while fitting demand
– Save money, optimal usage.
– Improve reliability and efficiency of grid
– Integration of green energy, reduction of CO2

More than 3.4 billion from US federal stimulus bill is targeted.
– Obama stimulus plan

One of hottest topic in research community
– But what are the problems from signal processing, communication
and networking points of view?
Are more easily
integrated into
power sys. Less
depend on fossil fuel
Gather, monitor the
usage so the supply more
efficiently and anticipate
challenging peaks
Realtime analysis,
Manage,
and
Smartplan,
Grid
forecast the energy
in-home
system
to management
meets the
tool needs
to track usage
Use sophisticated
comm. Technology
to find/fix problems
faster, enhancing
reliability
Connect grid to
charge overnight
when demand is low
Can generate own and
sellback excess energy
Supervisory Control and Data Acquisition Center

Real-time data acquisition
– Noisy analog measurements

Voltage, current, power flow
– Digital measurements

State estimation
– Maintain system in normal
state
– Fault detection
– Power flow optimization
– Supply vs. demand
SCADA TX data from/to Remote
Terminal Units (RTUs), the
substations in the grid
Privacy & Security Concern

More connections, more technology are linked to the obsolete
infrastructure.
– Add-on network technology: sensors and controls estimation
– More substations are automated/unmanned

Vulnerable to manipulate by third party
– Purposely blackout
– Financial gain
– Story of Enron
How to tackle this
issue at this moment?
Provide one example
next
Power System State Estimation Model

Transmitted active power from bus i to bus j
– High reactance over resistance ratio
– Linear approximation for small variance
– State vector
, measure noise e with covariance Ʃe
– Actual power flow measurement for m active power-flow branches
– Define the Jacobian matrix
– We have the linear approximation
– H is known to the power system but not known to the attackers
Bad Data Injection and Detection

State estimation from z

Bad data detection
– Residual vector
– Without attacker
where
– Bad data detection (with threshold )
without attacker:
with attacker:

otherwise
Stealth (unobservable) attack: z=Hx+c+e, where c=Hx
– Hypothesis test would fail in detecting the attacker, since the
control center believes that the true state is x + x.
Overview

Introduction to Smart Grid

Power System State Estimation Model

Bad Data Injection

Defender Mechanism
– Quickest Detection

Attacker Learning Scheme
– Independent Component Analysis

Future Work

A Few Topics in Smart Grid Communication

Conclusions

Quick View of Amigo Lab
Basics of Quickest Detection (QD)

Detect distribution changes of a sequence of observations

as quick as possible

with the constraint of false alarm or detection probability.
min [processing time]

s.t. Prob(true ≠ estimated) < ŋ
Classification
1. Bayesian framework:
 known prior information on probability
 SPRT (e.g. quality control, drug test, )
2. Non-Bayesian framework:
 unknown distribution and no prior
 CUSUM (e.g. spectrum sensing, abnormal detection )
QD System Model

Assuming Bayesian framework with non-stealthy attack
– the state variables are random with

The binary hypothesis test:

The distribution of measurement z under binary hyp: (differ
only in mean)

We want a detector
– False alarm and detection probabilities
Detection Model - NonBayesian

Non-Bayesian approach
– unknown prior probability, attacker statistic model

The unknown parameter exists
– in the post-change distribution and may changes over
the detection process.
– You do not know how attacker attacks.

Minimizing the worst-case effect via detection delay:
Detection
delay

Detection
time
Actual time of
active attack
We want to detect the intruder as soon as possible
while maintaining PD.
Multi-thread CUSUM Algorithm

CUSUM Statistic:
How about the
unknown?
where Likelihood ratio term of m measurements:

By recursion, CUSUM Statistic St at time t:
St = max[St-1 + Lt (Zt ), 0]

Average run length (ARL) for declaring attack with threshold h
Declare the attacker is existing!
Otherwise, continuous to the process.
Linear Solver for the Unknown

Rao test – asymptotically equivalent model of GLRT:

The linear unknown solver for m measurements:

Recursive CUSUM Statistic w/ linear unknown parameter solve:
– Modified CUSUM statistics
The unknown is no long
involved
m
ì
ü
é
T -1 T
-1 ù
St = max íSt-1 + åê( Zt SZ ) + SZ Zt ú, 0ý
ë
û þ
î
l=1
Simulation: Adaptive CUSUM algorithm

2 different detection tests: FAR: 1% and 0.1%

Active attack starts at time 5

Detection of attack at time 7 and 8, for different FARs
Markov Chain based Analytical Model

Divide statistic space into discrete states between 0 and threshold
– Obtain the transition probabilities
– Obtain expectation of detection delay, false alarm rate and missing
probability
Overview

Introduction to Smart Grid

Power System State Estimation Model

Bad Data Injection

Defender Mechanism
– Quickest Detection

Attacker Learning Scheme
– Independent Component Analysis

Future Work

A Few Topics in Smart Grid Communication

Conclusions

Quick View of Amigo Lab
Independent Component Analysis (ICA)

Linear Independent Component Analysis
– find a linear representation of the data so that components are
as statistically independent as possible.
– i.e., among the data, find how many independent sources.

Question for bad data injection:
– Without knowing H, the attacker can be caught.
– Could attacker launch stealthy attack to the system even
without knowledge about H?
– Using ICA, attacker could estimate H and consequently, lunch
an undetectable attack.
ICA Basics

A special case of blind source separation
u=Gv

u = [ui, i = 1, 2, … m]: observable vector

G = [gij, i = 1, 2, … m, j = 1, 2, … n]: mixing matrix
(unknown)

v = [vi, i = 1, 2, … n]: source vector (unknown)

Linear ICA implementation: FastICA from [Hyvärinen]
Stealth False Data Injection with ICA

Supposing that the noise is small, then we what to do the
mapping:
u=Gv
z=Hx

Problem: state vector x is highly correlated

Consider: x = A y, where
– A: constant matrix that can be estimated
– y: independent random vectors

Then we can apply Linear ICA on z = HA y
– We cannot know H, but we can know HA
– Stealthy attack: Z=Hx+HAy+e
Numerical Simulation Setting

Simulation setup
– 4-Bus test system, IEEE 14-Bus and 30-bus
– Matpower
Numerical Results

MSE of ICA inference (z-Gy) vs. the number of observations
(14-bus case).
Performance of the Attack
The PDF is the same w or w/o attacking.
So log likelihood is equal to 1– unable to detect
Overview

Introduction to Smart Grid

Power System State Estimation Model

Bad Data Injection

Defender Mechanism
– Quickest Detection

Attacker Learning Scheme
– Independent Component Analysis

Future Work

A Few Topics in Smart Grid Communication

Conclusions

Quick View of Amigo Lab
1. Distributed Smart Grid State Estimation

The deregulation has led to the creation of many regional
transmission organizations within a large interconnected power
system.

A distributed estimation and control is need .
– Distributed observability analysis
– Bad data detection

Challenges:
– Bottleneck and reliability problems with one coordination center.
– Need for wide area monitoring and control
– Convergence and optimality
Fully-Distributed State Estimation

With N substations/nodes
Local
observati
on matrix
Local Jacobian
matrix
Useful
information
to be
detected
Unknown
State
– By iteratively exchanging information with neighbors
– All local control center can achieve an unbiased consensus of
system-wide state estimation.
2. Optimality of Fault Detection Algorithm

Detecting the attack as an intermediate step towards obtaining a
reliable estimate about the injected false data
– Facilitates eliminating the disruptive effects of the false data

Joint estimation and detection problem
– Define an estimation performance measure
– Seek to the optimize it while ensuring satisfactory of the detection
performance
Performance
measurement
3. Manipulate Electricity Market
Example: Ex Post Market
Market that recalculate optimal points for generation and
consumption based on real-time data
I
Min :
St:

Generation Cost
*
C i ( Pg i   Pg i )
i 1
I

i 1
I
 Pg i    PL
i 1
 Pg i
 Fl
Power Balance
min
min
  Pg i   Pg i
*
  Fl   Fl
max
max
 i  1,..., I
 l  1,..., L
Generation & Transmission limits
[28]
4. PMU

PMU can measure voltage angle directly
– Defender: placement problem, no need to place nearby
– Attackers’ new strategy with existence of PMU
PMU
1
PMU
2
PMU
3
PMU
4
5
PMU
PMU
6
7
PMU
[29]
5. Game Theory Analysis
(attacker,
defender)
N
A
N
(0,0)
(b,-b)
D
(c,-c)
(-a,a)
a, b, c
How to formulate the game?
t
A Few Topics in Smart Grid Communications

Bad data injection

Demand side management
– Peak to average ratio
– Scheduling problem

Renewable energy
– The renewable energy is unreliable.
– Have to use diesel generators during shortage
– Not cheap and not green

PHEV
– routing, scheduling and resource allocation

Communication link effect on the smart grid
Conclusions


Bad data injection problem formulation
From defender point of view
– detect malicious bad data injection attack as quick as possible
– Adaptive CUSUM algorithm

From attacker point of view
– can estimate both the system topology and power states just by
observing the power flow measurements
– Independent component analysis algorithm to obtain information
– Once the information is at hand, malicious attacks can be launched
without triggering the detection system



Many possible future work
Edited book 2012 by Cambridge with E. Hossain and V. Poor.
Possible future collaboration
Overview of Wireless Amigo Lab

Lab Overview
– 7 Ph.D. students, 2 joint postdocs (with Rice and Princeton)
– supported by 5 NSF,1 DoD, and 1 Qatar grants

Current Concentration
– Game theoretical approach for wireless networking
– Compressive sensing and its application
– Smartgrid communication
– Bayesian nonparametric learning
– Security: trust management, belief network, gossip based Kalman
– Physical layer security
– Quickest detection
– Cognitive radio routing/security
– Sniffing: femto cell and cloud computing

USRP2 Implementation Testbed
Questions
Thank you for listening and supporting!