PaperlayerView exact revision on arXiv

Source-linked AI summary

Sparse Attack Construction and State Estimation in the Smart Grid: Centralized and Distributed Models

Mete Ozay, Inaki Esnaola, Fatos T. Yarman Vural, Sanjeev R. Kulkarni, H. Vincent Poor

arXiv:1502.04254v1cs.ITeess.SY

TL;DR

The paper addresses sparse false-data attack construction and state estimation in smart grids under centralized and distributed network structures. It proposes optimization-based attack and estimation models for different information and resource-sharing assumptions, and evaluates them in IEEE test systems. The experiments report comparable distributed estimation performance across centralized and massively distributed settings, while collective sparse attacks outperform distributed sparse attacks for constructing unobservable vectors.

  • Problem

    The paper addresses how to construct unobservable sparse data-injection attacks and estimate smart-grid states when network resources, measurements, and state variables are distributed.

  • Method

    The paper introduces centralized LASSO and selective attack constructions, distributed state-estimation methods, and distributed or collective sparse attack models solved through corresponding optimization procedures.

  • Results

    Collective Sparse Attacks perform better than Distributed Sparse Attacks for constructing unobservable attack vectors, while the two distributed state-estimation methods perform similarly across attack counts and test systems.

  • Takeaways & Limitations

    The proposed frameworks support sparse attack construction and state estimation across centralized, clustered, grouped, and massively distributed smart-grid settings.

  • Takeaways & Limitations

    The Distributed Sparse Attacks model assumes attackers process only local measurements to achieve consensus on attack vectors.

Abstract

from arXiv · show

New methods that exploit sparse structures arising in smart grid networks are proposed for the state estimation problem when data injection attacks are present. First, construction strategies for unobservable sparse data injection attacks on power grids are proposed for an attacker with access to all network information and nodes. Specifically, novel formulations for the optimization problem that provide a flexible design of the trade-off between performance and false alarm are proposed. In addition, the centralized case is extended to a distributed framework for both the estimation and attack problems. Different distributed scenarios are proposed depending on assumptions that lead to the spreading of the resources, network nodes and players. Consequently, for each of the presented frameworks a corresponding optimization problem is introduced jointly with an algorithm to solve it. The validity of the presented procedures in real settings is studied through extensive simulations in the IEEE test systems.

I. INTRODUCTION

The paper develops centralized and distributed sparse attack and state-estimation models for smart grids, motivated by network structure, limited information, and computational scale. It formulates unobservable attacks and distributed estimation procedures for detecting attacks.

  • I. INTRODUCTION: Power-network structure constrains the design of smart-grid attack and defense schemes, including centralized and distributed approaches.
  • I. INTRODUCTION: Centralized targeted attacks tamper with selected measurements to construct a specific false-data vector, while strategic attacks account for attacker-controlled and secure measurements.
  • I. INTRODUCTION: Distributed estimation is motivated by large-scale networks because computation can be spread across nodes without requiring every group to know the full network state.
  • I. INTRODUCTION: Distributed Sparse Attacks reach consensus from cluster-local measurements, whereas Collective Sparse Attacks apply shared attack vectors to grouped state variables with network-wide topology access.
  • I. INTRODUCTION: Distributed state estimation either groups measurements into clusters for processor-updated local estimates or assigns subsets of state variables to collaborating operators.
  • System model: Undetectable attacks satisfy a = Hc, placing the injected vector in H's column space, while state estimation uses weighted residual testing against a threshold.

B. Sparsity in the System

The paper exploits sparsity both in attack vectors and in temporal state changes, while accounting for measurement-matrix structure in IEEE test systems.

  • B. Sparsity in the System: Sparse attack construction limits tampering to at most k meters, with unobservable vectors guaranteed to exist when k ≥ N − D + 1.
  • B. Sparsity in the System: State estimation can recover only variables whose changes exceed a significance threshold, making the change vector nearly sparse over suitable time intervals.
  • B. Sparsity in the System: The nearly sparse state-change vector is recovered through compressed sensing with a regularization parameter γ.
  • B. Sparsity in the System: Several IEEE Jacobian matrices are rank deficient, while larger non-rank-deficient systems can still exhibit greater matrix sparsity ratios.

III. CENTRALIZED DATA SPARSE ATTACKS

Centralized sparse attack methods construct targeted false-data vectors by exploiting sparsity, secure measurements, and optimization formulations solved with ADMM.

  • III. CENTRALIZED DATA SPARSE ATTACKS: Targeted false-data attacks use LASSO and regressor-selection formulations to construct sparse vectors for specified state-variable attack targets.
  • III. CENTRALIZED DATA SPARSE ATTACKS: The construction decomposes targeted and off-target contributions, then solves for a sparse attack vector through a projected measurement relation.
  • III. CENTRALIZED DATA SPARSE ATTACKS: LASSO balances sparsity through the ℓ1 regularizer λ and detection risk through a residual cost, and its optimization is solved using ADMM.
  • III. CENTRALIZED DATA SPARSE ATTACKS: The LASSO procedure initializes ADMM variables, performs ridge regression and soft thresholding, and iterates until stopping criteria or a maximum iteration count is reached.
  • III. CENTRALIZED DATA SPARSE ATTACKS: Selective attacks replace implicit sparsity control with the explicit constraint ∥a∥0 ≤ k.

2) Selective Targeted Attacks:

Selective targeted attacks solve sparse attack construction with hard thresholding, while strategic sparse attacks enforce zero impact on secure measurements through a Jacobian decomposition.

  • 2) Selective Targeted Attacks:: Selective optimization uses hard thresholding to retain the k largest-magnitude elements and set the remaining elements to zero.
  • B. Strategic Sparse Attacks: The strategic sparse attack algorithms compute c using LASSO and regressor-selection formulations adapted for the strategic attack model.
  • B. Strategic Sparse Attacks: Strategic sparse attacks decompose the Jacobian into attacked and secure measurement rows and require the secure-measurement contribution H_Sc to equal zero.
  • B. Strategic Sparse Attacks: The strategic attack formulation searches for a state perturbation c consistent with secure measurements while constructing the attack through the accessible meters.

1) Strategic Sparse Attacks with LASSO:

Strategic sparse attacks formulate attack construction under secure-measurement constraints and solve the resulting sparsity-controlled problems with ADMM.

  • 1) Strategic Sparse Attacks with LASSO:: Strategic attacks assume control over only selected measurements while excluding secure measurements from attack construction.
  • 1) Strategic Sparse Attacks with LASSO:: The sparse optimization is reformulated as a LASSO problem and then as an ADMM problem solved repeatedly for each attack-vector component.
  • 1) Strategic Sparse Attacks with LASSO:: The formulation relaxes H_Sc = 0 and imposes at most k nonzero attack-vector elements, using hard thresholding for the constrained solution.
  • 1) Strategic Sparse Attacks with LASSO:: The centralized algorithm has complexity Υ1 ∈ O(t′α^3), with α determined by the targeted or strategic attack setting.For strategic attacks, repeating the procedure D times increases complexity to O(t′Dα^3).

IV. DISTRIBUTED AND COLLABORATIVE SPARSE STATE VECTOR ESTIMATION

The paper introduces distributed sparse state estimation that partitions measurements into clusters, computes local estimates, and enforces a shared global state through sparse consensus optimization.

  • IV. DISTRIBUTED AND COLLABORATIVE SPARSE STATE VECTOR ESTIMATION: Distributed Sparse State Vector Estimation uses an instance distributed LASSO algorithm to estimate state vectors under attacked network measurements.
  • IV. DISTRIBUTED AND COLLABORATIVE SPARSE STATE VECTOR ESTIMATION: Measurements and attack vectors are partitioned into G cluster blocks, enabling local processing across network operators.
  • IV. DISTRIBUTED AND COLLABORATIVE SPARSE STATE VECTOR ESTIMATION: The distributed optimization uses β as a global variable and a consensus constraint so operators agree on the estimated state vector.
  • IV. DISTRIBUTED AND COLLABORATIVE SPARSE STATE VECTOR ESTIMATION: The optimization problems are solved with an ADMM implementation using cluster sets and state measurements as inputs.
  • IV. DISTRIBUTED AND COLLABORATIVE SPARSE STATE VECTOR ESTIMATION: Each cluster computes a Tikhonov-regularized local estimate, after which local estimates update a global variable and are broadcast for dual-variable updates.

B. Collaborative Sparse State Vector Estimation

Collaborative sparse state estimation distributes groups of state variables across clusters while coordinating their estimates through a shared optimization variable and ADMM.

  • B. Collaborative Sparse State Vector Estimation: Operators may know the full topology and Jacobian matrix while observing different groups of state-vector variables across clusters.
  • B. Collaborative Sparse State Vector Estimation: The observation model represents measurements as a noisy function of a cluster-specific state vector and a Jacobian submatrix selecting that cluster’s variables.
  • B. Collaborative Sparse State Vector Estimation: Introducing a shared optimization variable v makes the collaborative formulation equivalent to the distributed problem and enables an ADMM solution.
  • B. Collaborative Sparse State Vector Estimation: Collaborative Estimation via ADMM is specified with the projection matrix, state measurements, and cluster set as inputs.

V. DISTRIBUTED AND COLLECTIVE SPARSE ATTACKS

The paper extends sparse attack construction to distributed settings with separate models for locally informed clustered attackers and topology-informed attacks targeting distributed state variables.

  • V. DISTRIBUTED AND COLLECTIVE SPARSE ATTACKS: Two distributed attack models reformulate false-data injection as distributed sparse optimization after redefining measurement and attack-vector structure.
  • V. DISTRIBUTED AND COLLECTIVE SPARSE ATTACKS: Distributed Sparse Attacks let attackers in different clusters construct vectors using only local measurements, while Collective Sparse Attacks assume knowledge of the full topology and Jacobian matrix.
  • V. DISTRIBUTED AND COLLECTIVE SPARSE ATTACKS: The distributed attack strategies can also serve as design patterns for distributed sparse targeted and strategic false-data injection attacks.
  • V. DISTRIBUTED AND COLLECTIVE SPARSE ATTACKS: The linear sparse attack model seeks c such that a = Hc and uses an ℓ1 relaxation to estimate a sparse injection vector.
  • V. DISTRIBUTED AND COLLECTIVE SPARSE ATTACKS: Separability of the loss function yields a distributed formulation with local variables φ and ℓ1 regularization, solvable using the same ADMM-based approach as distributed estimation.

B. Collective Sparse Attacks

Collective Sparse Attacks distribute attack-vector variables across clusters while using shared network information and iterative optimization to coordinate local solutions.

  • Collective attack model: Collective Sparse Attacks assume attackers know the full network topology and Jacobian matrix H while targeting groups of state-vector variables.Each attacker computes an injection vector from a Jacobian measurement submatrix associated with its cluster.
  • Optimization formulation: The collective formulation introduces auxiliary optimization variables and equality constraints linking each local injection vector c_i to its corresponding estimate ψ̂_i.The resulting problem is equivalent to an earlier formulation, so Algorithm 3 solves it.
  • Distributed optimization: The distributed framework uses measurement-distributed and attribute-distributed optimization, depending on whether measurements or state and attack variables are partitioned.Local solutions are computed within clusters under the corresponding partitioning assumption.
  • Complexity: Algorithm 2 similarly has per-cluster cost O(α_i^3), with α_i = min(N_i,D), and communication raises its complexity by Υ_g + Υ_b.A central processor waits for local estimates from all clusters, making the maximum cluster cost decisive.
  • Complexity: Ignoring communication, Algorithm 3 has parallel local least-squares updates with per-cluster cost O(α_i^3), where α_i = min(N,D_i).The overall computational cost is governed by the slowest cluster and increases by Υ_g + Υ_b for gathering and broadcasting.

VII. NUMERICAL RESULTS

Numerical experiments on IEEE test systems evaluate sparse attack construction probabilities under varying attack sparsity and compare LASSO-based with selective attack methods.

  • Evaluation measures: The simulations measure correct attacked-variable and secure-variable construction probabilities, with incorrect-construction probabilities obtained from their complements.The attack model randomly generates a k-sparse vector using uniformly selected nonzero indices and Gaussian amplitudes matched to z.
  • Experimental setup: Dynamic λ computation makes constructed false-data vectors exhibit similar probabilities across test systems and sparsity levels while estimating attack sparsity.The experiments vary the number of attack variables across IEEE 9-, 30-, 57-, and 118-bus systems, using 100 realizations per data point.
  • Centralized attack comparisons: TLA attacked-variable construction probabilities range from 0.5 to 0.7, while secure-variable probabilities range from 0.3 to 0.5 in Figure 1.These values are reported for the corresponding Figure 1 panels as the number of attack variables varies.
  • Centralized attack comparisons: Selective Attacks show smooth changes in false-data construction probabilities as the number of attack variables varies because their sparsity is controllable.By contrast, TLA probabilities do not change smoothly because λ is dynamically computed for each realization and test system.
  • Centralized attack comparisons: LASSO and selective attacks produce similar attack-vector solutions when the LASSO regularization parameter is optimized, with SLA and SSA showing similar probability trends.Attacked-variable construction probabilities increase similarly, while secure-variable probabilities decrease similarly across the two methods.

B. Results for Attack Detection using Distributed Sparse State Vector Estimation

The proposed distributed state-estimation methods are evaluated for attack detection on IEEE test systems using precision, recall, and accuracy. Their performance remains similar across cluster counts, with small loss relative to centralized strategies in the simulated settings.

  • The attack-detection procedure estimates state vectors first, then computes measurement residuals and flags observation i when its residual exceeds τ.Attacked measurements are retained rather than removed at each iteration, avoiding dataset updates and additional parameter re-estimation costs.
  • Precision, recall, and accuracy are used to evaluate how well attacked measurements are detected.Precision concerns correctly detecting specific attacks, while true-positive, true-negative, false-positive, and false-negative counts define the metrics.
  • The distributed and collective algorithms have similar performance across different cluster counts, with small optimality loss relative to centralized strategies.Figure 3 compares varying G values for the IEEE 57-bus system, including extreme distributed-processing cases.
  • For the IEEE-57-bus system, precision fluctuates at low k/N and stabilizes around 0.9 as k/N increases.The corresponding accuracy and recall curves increase with k/N, with slightly smaller slopes than for the IEEE-9-bus system.

C. Results for Distributed and Collective Sparse Attacks

The paper evaluates distributed and collective sparse-attack constructions alongside centralized and distributed estimation methods. Results show strong attack-injection performance, similar estimation behavior across centralized and distributed settings, and better collective than distributed attack construction in the reported experiments.

  • The proposed methods successfully inject data with high probability across a large range of sparsity levels in IEEE-57-bus distributed attack experiments.The reported attack-detection probability decreases and error increases as G decreases, while smaller G makes performance less sensitive to regularization-parameter tuning.
  • The paper introduces LASSO Attacks and Selective Attacks for centralized false-data and attack-vector construction under Targeted and Strategic attack models.Selective Attacks explicitly control the sparsity of attack vectors by specifying attacked and secure variables.
  • Centralized and massively distributed estimation achieve similar performance, while accuracy and precision decrease as system size increases.The experiments report little change in performance as the number of clusters increases, including G = 1 and the extreme distributed cases.
  • Collective Sparse Attacks perform better than Distributed Sparse Attacks for constructing unobservable attack vectors in the experiments.For larger systems, higher G values can outperform smaller G values, partly because parameter estimation challenges affect the proposed methods.
Loading 1502.04254v1…