a deep learning-based cyber- physical strategy to mitigate false … · 2016. 4. 21. · deep...

31
A Deep Learning-Based Cyber- Physical Strategy to Mitigate False Data Injection Attack in Smart Grids Jin Wei and Gihan J. Mendis Department of Electrical & Computer Engineering 04/12/2016

Upload: others

Post on 08-Oct-2020

1 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: A Deep Learning-Based Cyber- Physical Strategy to Mitigate False … · 2016. 4. 21. · Deep Learning-Based Cyber-Physical Strategy • The PDC works as an aggregator in the intra-cluster

A Deep Learning-Based Cyber-Physical Strategy to MitigateFalse Data Injection Attack in Smart Grids

Jin Wei and Gihan J. Mendis

Department of Electrical & Computer Engineering

04/12/2016

Page 2: A Deep Learning-Based Cyber- Physical Strategy to Mitigate False … · 2016. 4. 21. · Deep Learning-Based Cyber-Physical Strategy • The PDC works as an aggregator in the intra-cluster

Outline

• Introduction

• Problem Setting

• Attack-Mitigation Model

• Deep Learning-Based Cyber-Physical Strategy

• Simulations

• Conclusions

2

Page 3: A Deep Learning-Based Cyber- Physical Strategy to Mitigate False … · 2016. 4. 21. · Deep Learning-Based Cyber-Physical Strategy • The PDC works as an aggregator in the intra-cluster

Introduction

• The attacks targeting availability, integrity, and confidentiality are identified as the major security issues for smart grids.

• We focus on one typical data integrity attack, false data injection attack.

3

Page 4: A Deep Learning-Based Cyber- Physical Strategy to Mitigate False … · 2016. 4. 21. · Deep Learning-Based Cyber-Physical Strategy • The PDC works as an aggregator in the intra-cluster

Introduction

• The tremendous amount of real-time data, such as PMU data, demands the appropriate data analysis and control techniques.

• Deep Belief Network (DBN) is coming to play a key role in providing big data predictive analytics solutions.

4

Page 5: A Deep Learning-Based Cyber- Physical Strategy to Mitigate False … · 2016. 4. 21. · Deep Learning-Based Cyber-Physical Strategy • The PDC works as an aggregator in the intra-cluster

Problem Setting• A two-tier hierarchical cyber-physical multi-agent control

framework for modeling the transient stability problem.

Agent 9Agent 7

Agent 3

5

Page 6: A Deep Learning-Based Cyber- Physical Strategy to Mitigate False … · 2016. 4. 21. · Deep Learning-Based Cyber-Physical Strategy • The PDC works as an aggregator in the intra-cluster

Problem Setting

• A two-tier hierarchical cyber-physical multi-agent control framework for modeling the transient stability problem.

– Inter-cluster (Tier-1): lead agents achieve frequency synchronization via cyber-physical couplings.

– Intra-cluster (Tier-2): multiple secondary agents achieve synchronization via strong physical couplings with a stabilized lead agent.

• Only lead agent PMU information is needed to ensure transient system stabilization in the face of a disturbance.

6

Page 7: A Deep Learning-Based Cyber- Physical Strategy to Mitigate False … · 2016. 4. 21. · Deep Learning-Based Cyber-Physical Strategy • The PDC works as an aggregator in the intra-cluster

Problem Setting

• A two-tier hierarchical cyber-physical multi-agent control framework for modeling the transient stability problem.

– Inter-cluster (Tier-1): lead agents achieve frequency synchronization via cyber-physical couplings.

– Intra-cluster (Tier-2): multiple secondary agents achieve synchronization via strong physical couplings with a stabilized lead agent.

• Only lead agent PMU information is needed to ensure transient system stabilization in the face of a disturbance.

How to defend against the false data injection attack on the lead agent PMUs?

7

Page 8: A Deep Learning-Based Cyber- Physical Strategy to Mitigate False … · 2016. 4. 21. · Deep Learning-Based Cyber-Physical Strategy • The PDC works as an aggregator in the intra-cluster

Attack-Mitigation Model

8

Page 9: A Deep Learning-Based Cyber- Physical Strategy to Mitigate False … · 2016. 4. 21. · Deep Learning-Based Cyber-Physical Strategy • The PDC works as an aggregator in the intra-cluster

Deep Learning-Based Cyber-Physical Strategy

• In the hierarchical framework, the states of the secondary agents can be treated as “noisy” version of those of the lead agents.

• The PMU data from the lead agents can be validated using the PMU data from the secondary agents.

9

Page 10: A Deep Learning-Based Cyber- Physical Strategy to Mitigate False … · 2016. 4. 21. · Deep Learning-Based Cyber-Physical Strategy • The PDC works as an aggregator in the intra-cluster

Deep Learning-Based Cyber-Physical Strategy

• The PDC works as an aggregator in the intra-cluster LAN to verify the trustworthiness of the lead agent’s PMU data.

10

Page 11: A Deep Learning-Based Cyber- Physical Strategy to Mitigate False … · 2016. 4. 21. · Deep Learning-Based Cyber-Physical Strategy • The PDC works as an aggregator in the intra-cluster

Deep Learning-Based Cyber-Physical Strategy

• The PDC works as an aggregator in the intra-cluster LAN to verify the trustworthiness of the lead agent’s PMU data.

– The PDC probes the PMU data from all of the secondary agents at a verification rate for learning the features of the PMU data of the cluster of agents and thus predicting the behavior pattern of the PMU data of the associated lead agent.

• Using Deep-learning Technique

– The PDC also probes the lead agent’s data at the same rate and measure the trustworthiness of a lead agent’s data by using the predicted behavior pattern.

11

Page 12: A Deep Learning-Based Cyber- Physical Strategy to Mitigate False … · 2016. 4. 21. · Deep Learning-Based Cyber-Physical Strategy • The PDC works as an aggregator in the intra-cluster

Deep Learning-Based Cyber-Physical Strategy

12

Page 13: A Deep Learning-Based Cyber- Physical Strategy to Mitigate False … · 2016. 4. 21. · Deep Learning-Based Cyber-Physical Strategy • The PDC works as an aggregator in the intra-cluster

Deep Learning-Based Cyber-Physical Strategy

13

No. Input

I

II

Page 14: A Deep Learning-Based Cyber- Physical Strategy to Mitigate False … · 2016. 4. 21. · Deep Learning-Based Cyber-Physical Strategy • The PDC works as an aggregator in the intra-cluster

Deep Learning-Based Cyber-Physical Strategy

14

Hid

den

La

yer 1

Hid

den

La

yer

3

Hid

den

La

yer

4

Hid

den

La

yer 2

)1(i

tx)1(1

th

i

tx1

th2

th

)2(1

th

)3(1

th

)1(2

th

)2(2

th

)3(2

th

)1(3

th

)2(3

th)1(4

th

3

th

)2(i

tx

)3(i

tx

Fig. 1: Conditional Deep Belief Network

Page 15: A Deep Learning-Based Cyber- Physical Strategy to Mitigate False … · 2016. 4. 21. · Deep Learning-Based Cyber-Physical Strategy • The PDC works as an aggregator in the intra-cluster

Deep Learning-Based Cyber-Physical Strategy

• Considering that the 3-dimensional input data have real values, both of the CDBNs are trained by stacking one conditional GBRBM with three conditional BRBMs.

– For conditional GBRBM: Energy function.

15

where

Page 16: A Deep Learning-Based Cyber- Physical Strategy to Mitigate False … · 2016. 4. 21. · Deep Learning-Based Cyber-Physical Strategy • The PDC works as an aggregator in the intra-cluster

Deep Learning-Based Cyber-Physical Strategy

• Considering that the 3-dimensional input data have real values, both of the CDBNs are trained by stacking one conditional GBRBM with three conditional BRBMs.

– For conditional GBRBM: conditional probability distributions.

16

where

Page 17: A Deep Learning-Based Cyber- Physical Strategy to Mitigate False … · 2016. 4. 21. · Deep Learning-Based Cyber-Physical Strategy • The PDC works as an aggregator in the intra-cluster

Deep Learning-Based Cyber-Physical Strategy

• Considering that the 3-dimensional input data have real values, both of the CDBNs are trained by stacking one conditional GBRBM with three conditional BRBMs.

– For conditional BRBM: Energy function.

17

where

Page 18: A Deep Learning-Based Cyber- Physical Strategy to Mitigate False … · 2016. 4. 21. · Deep Learning-Based Cyber-Physical Strategy • The PDC works as an aggregator in the intra-cluster

Deep Learning-Based Cyber-Physical Strategy

• Considering that the 3-dimensional input data have real values, both of the CDBNs are trained by stacking one conditional GBRBM with three conditional BRBMs.

– For conditional BRBM: conditional probability distributions.

18

where

Page 19: A Deep Learning-Based Cyber- Physical Strategy to Mitigate False … · 2016. 4. 21. · Deep Learning-Based Cyber-Physical Strategy • The PDC works as an aggregator in the intra-cluster

Deep Learning-Based Cyber-Physical Strategy

19

Page 20: A Deep Learning-Based Cyber- Physical Strategy to Mitigate False … · 2016. 4. 21. · Deep Learning-Based Cyber-Physical Strategy • The PDC works as an aggregator in the intra-cluster

Deep Learning-Based Cyber-Physical Strategy

20

Page 21: A Deep Learning-Based Cyber- Physical Strategy to Mitigate False … · 2016. 4. 21. · Deep Learning-Based Cyber-Physical Strategy • The PDC works as an aggregator in the intra-cluster

Deep Learning-Based Cyber-Physical Strategy

21

Page 22: A Deep Learning-Based Cyber- Physical Strategy to Mitigate False … · 2016. 4. 21. · Deep Learning-Based Cyber-Physical Strategy • The PDC works as an aggregator in the intra-cluster

Simulations

Fig. 2: New England 39-bus power system

22

Page 23: A Deep Learning-Based Cyber- Physical Strategy to Mitigate False … · 2016. 4. 21. · Deep Learning-Based Cyber-Physical Strategy • The PDC works as an aggregator in the intra-cluster

Simulations

23

Fig. 3: Normalized rotor frequencies and phase angles versus time without active control and proposed deep learning-based cyber-physical strategy

Page 24: A Deep Learning-Based Cyber- Physical Strategy to Mitigate False … · 2016. 4. 21. · Deep Learning-Based Cyber-Physical Strategy • The PDC works as an aggregator in the intra-cluster

Simulations

24

4

Page 25: A Deep Learning-Based Cyber- Physical Strategy to Mitigate False … · 2016. 4. 21. · Deep Learning-Based Cyber-Physical Strategy • The PDC works as an aggregator in the intra-cluster

Simulations

25

𝑈[−360o, 360o]

Page 26: A Deep Learning-Based Cyber- Physical Strategy to Mitigate False … · 2016. 4. 21. · Deep Learning-Based Cyber-Physical Strategy • The PDC works as an aggregator in the intra-cluster

Simulations

26

Fig. 4: Normalized rotor frequencies and phase angles versus time withactive control and without the proposed deep learning-based cyber-physical strategy in presence of random attack.

Page 27: A Deep Learning-Based Cyber- Physical Strategy to Mitigate False … · 2016. 4. 21. · Deep Learning-Based Cyber-Physical Strategy • The PDC works as an aggregator in the intra-cluster

Simulations

27

Page 28: A Deep Learning-Based Cyber- Physical Strategy to Mitigate False … · 2016. 4. 21. · Deep Learning-Based Cyber-Physical Strategy • The PDC works as an aggregator in the intra-cluster

Simulations

28

Page 29: A Deep Learning-Based Cyber- Physical Strategy to Mitigate False … · 2016. 4. 21. · Deep Learning-Based Cyber-Physical Strategy • The PDC works as an aggregator in the intra-cluster

Simulations

29

Fig. 5: Normalized rotor frequencies and phase angles versus time withactive control and the proposed deep learning-based cyber-physical strategy in presence of random attack.

Page 30: A Deep Learning-Based Cyber- Physical Strategy to Mitigate False … · 2016. 4. 21. · Deep Learning-Based Cyber-Physical Strategy • The PDC works as an aggregator in the intra-cluster

Conclusions• We proposed the a deep learning-based approach to detect

and defend against the potential false data injection attack on the critical data for the application of maintaining the transient stability of real-time WAMS.

• Future work will examine a generalized class of threat models for which the approach is able to identify data corruption.

30

Page 31: A Deep Learning-Based Cyber- Physical Strategy to Mitigate False … · 2016. 4. 21. · Deep Learning-Based Cyber-Physical Strategy • The PDC works as an aggregator in the intra-cluster

Questions

31