Source-linked AI summary
Attack-Resiliency Analytics for Wide-Area Control Systems in Smart Grids
Mohammad Zakaria Haider, Prabin Mali, Nur Imtiazul Haque, Muhammad Nadeem, Sumit Paudyal, Mohammad Ashiqur Rahman
TL;DR
Wide-area control systems can be exposed to stealthy false data injection attacks through their networked measurements. The paper models coupled grid-control dynamics and optimizes stealthy attacks while comparing static and learned detector constraints. Wide-area access amplifies attack severity, while learned boundaries reduce severity but do not eliminate relay-relevant excursions.
Problem
Stealthy false data injection can manipulate measurements and actuate protection relays, while inter-area oscillations require wide-area coordination to damp them.
Source: §I · p. 1, C. Anomaly Detection Model · p. 3Method
The framework formulates stealthy attack optimization as a mixed-integer linear program over coupled control dynamics, with replaceable static BDD or learned detector constraints calibrated to a common benign false-positive rate.
Source: §I · p. 2, B. Data-Driven ADM Constraints · p. 4Results
Wide-area access yields 3.3× and 8.1× the benign objective on the 39- and 118-bus systems; learned constraints reduce the objective by 15.5–55.0% but every configuration retains an attack that reaches the 60.5 Hz crossing.
Source: §VI · p. 9Takeaways & Limitations
Security priority should reflect compromise impact rather than benign contribution alone, and learned-boundary residual severity depends more on boundary width than detector family.
Source: §VI · p. 9The framework models immediate attack impacts but excludes cascading failures and communication latency.
Source: §VI · p. 9
Abstract
from arXiv · showhide
Wide-area monitoring, protection, and control (WAMPAC) systems damp inter-area oscillations in large interconnected grids, but their reliance on synchronized PMU measurements carried over wide-area networks exposes them to false data injection (FDI) attacks. This paper presents an attack-resiliency analytics framework that formally models the coupled dynamics of the wide-area damping loop, automatic generation control, and the governor and excitation systems, and formulates the optimal stealthy FDI attack as a mixed-integer linear program (MILP). The anomaly detection model (ADM) enters as a replaceable constraint set: either a static bad-data detection (BDD) rule or a boundary learned from benign operation calibrated to a common false-positive rate. On the IEEE 39 and 118 bus systems, with reference dynamics validated on an OPAL-RT hardware-in-the-loop testbed, the optimal attack with wide-area access attains 3.3 and 8.1 times the benign objective and reaches a 0.5Hz frequency excursion more than twice as fast as an attack confined to automatic generation control. Learned boundaries reduce the attack objective by 15.5 55.0% and prevent over-frequency relay trips in all configurations, with feasible stealthy attacks remaining in every case, and residual risk tracks the width of the learned boundary rather than the detector family.
I. INTRODUCTION
WAMPAC coordinates damping across interconnected grids but depends on networked PMU telemetry that can be manipulated. This work addresses gaps in prior isolated-loop attack studies by modeling coupled controls and comparing static and learned detectors.
I. INTRODUCTION
- WAMPAC uses synchronized PMU data for coordinated real-time control, but networked telemetry exposes it to FDI attacks that can manipulate measurements to evade anomaly detection.Inter-area oscillations require wide-area coordination, unlike local oscillation modes that local controllers can damp.Source: §I · p. 1
II. RELATED WORK
- Prior studies largely isolate control loops and static BDD rules, leaving coupled control dynamics and comparisons with learned detectors unaddressed.The paper evaluates the framework on two test systems of markedly different scales.Source: §I · p. 1, §II · p. 2
I. INTRODUCTION
- The framework models coupled WAMPAC, AGC, governor, excitation, and generator dynamics and maximizes cumulative frequency and voltage excursions under operational, access, and detector-consistency constraints.The resulting attack optimization is formulated as a MILP.Source: §I · pp. 1–2
- Learned stealth bounds are derived from benign PMU data and calibrated to a common false-positive rate for comparison with static BDD rules.The same attack problem is solved once per detector boundary.Source: §I · p. 2
- With wide-area access, optimal attacks attain 3.3× and 8.1× the benign objective on the 39- and 118-bus systems and reach a 0.5 Hz excursion more than twice as fast as AGC-only attacks.The reported orderings hold across both systems, with reference dynamics validated on an OPAL-RT hardware-in-the-loop testbed.Source: §I · p. 2
III. TECHNICAL OVERVIEW OF WAMPAC
The WAMPAC model connects a wide-area damping signal to local excitation control, while AGC and governor dynamics regulate generation and swing equations govern generator response.
A. System Dynamics
- WAMPAC uses synchronized PMU-based state estimation and a PID controller driven by inter-area speed deviations to generate a global control signal integrated with local PSS.The wide-area signal supplements standard PSS output and modifies generator excitation.Source: A. System Dynamics · p. 3
- The AVR uses the combined signal to adjust terminal voltage, and its closed-loop error responds to the modified reference.The stated role includes maintaining rotor-angle stability amid power-flow variations.Source: A. System Dynamics · p. 3
- AGC regulates frequency and scheduled tie-line flows by integrating area control error into a mechanical-power reference that the steam governor maps to mechanical power.The model uses a standard TGOV1 steam governor.Source: A. System Dynamics · p. 3
- Generator frequency behavior is governed by swing equations coupling mechanical input and electrical output, while normal wide-area operation routes corrections until frequency and angle deviations are neutralized.The stated equilibrium conditions include zero local frequency and angle deviations and zero inter-area deviations.Source: A. System Dynamics · p. 3
B. Protection Relays
The protection-relay model defines safe frequency deviation limits and treats excursions beyond them as relay-triggering conditions.
B. Protection Relays
- Frequency deviations outside [−0.5, 0.5] Hz trigger relays, while allowable voltage deviations are ±10% of nominal.The stated safe frequency range follows; governors correct minor deviations of ±0.02 to ±0.05 Hz.Source: B. Protection Relays · p. 3
C. Anomaly Detection Model
The framework treats anomaly detection as a replaceable constraint and formulates stealthy attacks against the accepted per-cycle changes while maintaining grid operating limits.
C. Anomaly Detection Model
- Static BDD alarms when the state-estimation residual exceeds a threshold, while the framework also tests learned boundaries calibrated to the same benign false-positive rate.The four data-driven detector families use per-cycle rates of change, and each channel’s learned boundary is converted into a scalar stealth constraint.Source: C. Anomaly Detection Model · p. 3
IV. FORMAL MODELING OF OPTIMAL FDI ATTACK
- The attack is formulated as a MILP that optimizes measurement manipulations to disrupt estimation and control while evading detection.The nonlinear WAMPAC dynamics are discretized with Backward Euler, and the program is solved with Gurobi.Source: §IV · p. 3
A. ADM Constraints for Controllers
- ADM constraints bound each compromised channel’s per-cycle change by its accepted threshold, while operational limits and ACE constraints restrict system behavior.The formulation uses continuous stealth slack variables; binary variables select accessible controllers and linearize absolute values in the objective.Source: A. ADM Constraints for Controllers · pp. 3–4
B. Data-Driven ADM Constraints
The framework derives detector-specific stealth bounds from benign operation and treats the anomaly detector as a replaceable parameter in a shared attack optimization. Learned bounds are calibrated at a common benign false-positive rate and can adapt over time.
B. Data-Driven ADM Constraints
- A common benign false-positive rate makes boundaries from different detector models directly comparable within the same attack optimization.The detector changes the constraint parameter, not the optimization problem.Source: B. Data-Driven ADM Constraints · p. 4
- Each learned bound is the smallest per-cycle excursion that the fitted detector still reports as anomalous.The detector is fitted to per-controller rates of change collected from the benign run.Source: B. Data-Driven ADM Constraints · p. 4
- A benign-trajectory margin prevents uncompromised controllers from violating their own learned ADM constraints and making the attack problem spuriously infeasible.The bound includes κ max Dχ with κ = 1.05.Source: B. Data-Driven ADM Constraints · p. 4
- Sliding-window recalibration produces piecewise-constant thresholds that widen during disturbances and tighten in steady state.The evaluation uses fixed-length windows of 10 TS.Source: B. Data-Driven ADM Constraints · p. 4
- Static BDD is the special case of a fixed detector and threshold, while learned detectors derive their bounds from benign data.The static thresholds are fixed at commissioning from engineering margins.Source: B. Data-Driven ADM Constraints · p. 4
C. Attack Technique
The attack technique injects malicious vectors into AGC, AVR, and WAMPAC measurement streams, producing compromised measurements. Its objective is to drive frequency and voltage away from nominal and toward relay thresholds.
C. Attack Technique
- FDI vectors target AGC frequency and AVR voltage measurements, while a separate payload targets PMU streams consumed by WAMPAC.The resulting compromised measurements are expressed by the FDI mapping.Source: C. Attack Technique · p. 4
- The described execution first passively intercepts unencrypted IEEE C37.118 streams to learn benign transients, then overwrites telemetry in real time with synthesized vectors.The passage gives man-in-the-middle techniques such as ARP poisoning and packet crafting as examples.Source: C. Attack Technique · p. 4
D. Attacker Goal
- The attacker maximizes cumulative absolute deviation of grid frequency and voltage from their nominal steady states to breach protective-relay limits.The objective is sign-agnostic, so the attack may drive the system toward under- or over-frequency.Source: D. Attacker Goal · pp. 4–5
E. Adversarial Attributes
The attack surface is limited by channel-specific network access and an accessibility budget, while the attack strategy respects per-cycle detector bounds. The analysis also imposes attacker-knowledge, measurement-access, and protected-control-signal assumptions.
E. Adversarial Attributes
- Accessibility is represented by budgeted boolean vectors for AGC, AVR, and WAMPAC channels; measurements at uncompromised buses remain immutable.The vectors encode whether the attacker can reach each channel at each bus.Source: E. Adversarial Attributes · p. 5
- On the 118-bus system, fixing the accessible set before optimization restricts controller selection, so reported severities are lower bounds under the same budget.The wide-area set is the only one requiring selection because βagc = 10 already covers every governor-equipped unit.Source: E. Adversarial Attributes · p. 5
F. Attack Strategies and ADM Evasion
- The attacker uses a trajectory-wide sequence of bounded increments that mimics legitimate load transients rather than a single disruptive injection.The strategy exploits the margin between per-cycle ADM tolerance and the cumulative excursion needed to reach relay thresholds.Source: F. Attack Strategies and ADM Evasion · p. 5
G. Attack Assumptions
- The attacker is assumed to know control and ADM properties and generator parameters, modify accessible PMU/IED measurements while control signals stay protected, and inject throughout the horizon or within a bounded window.Source: G. Attack Assumptions · p. 5
V. EVALUATION
The evaluation uses IEEE 39-bus and 118-bus models, with reference dynamics tested on an OPAL-RT hardware-in-the-loop platform. Baseline results show that WAMPAC reduces frequency oscillations, with similar responses observed on the testbed.
A. Hardware-in-the-Loop Setup
- The evaluation covers IEEE 39-bus and 118-bus systems, with real-time phasor-data handling and reference dynamics assessed on an OPAL-RT hardware-in-the-loop testbed.The 39-bus model has 10 synchronous generators and three control areas; the 118-bus system has 54 generators across three areas.Source: A. Hardware-in-the-Loop Setup · p. 5
- The canonical setup uses an 800 MW load disturbance, a 1/60 s time step, and a 600–1200-step optimization horizon.The disturbance is applied for about 3.3 s, beginning at TS 125.Source: A. Hardware-in-the-Loop Setup · p. 5
B. Baseline Validation
- WAMPAC reduced peak-to-peak frequency oscillations by approximately 10%, and OPAL-RT responses were similar to the model despite minor power-flow discrepancies.Figures 2 and 3 show responses without and with WAMPAC; Fig. 4 compares the linear optimizer with OPAL-RT.Source: B. Baseline Validation · p. 6, Figure 2 · p. 6, Figure 3 · p. 6
C. Impact of Continuous and Timebound FDI Attack
Attacks on AGC destabilize frequency, while attacks on WAMPAC disrupt wide-area damping and increase settling time. Adding access to AVR and WAMPAC lets an attacker reach a 0.5 Hz frequency deviation faster than with AGC access alone.
C. Impact of Continuous and Timebound FDI Attack
- A continuous AGC measurement attack caused frequency fluctuations up to 0.25 Hz within the eighth AGC cycle; a time-bound attack caused variations up to 0.15 Hz before stability returned.Similar deviation patterns were observed in voltage, rotor angles, and tie-line power flows.Source: C. Impact of Continuous and Timebound… · p. 6
D. Analysis of Optimal FDI Attack on WAMPAC
- An ongoing WAMPAC attack disrupted damping control and made the system take 30% longer to reach steady state than the response with WAMPAC integrated.Fig. 7 presents the response during the WAMPAC attack; the comparison baseline is Fig. 3.Source: D. Analysis of Optimal FDI Attack on… · p. 7
E. Evaluation of Impact Based on Accessibility
- Adding AVR and then WAMPAC access reduced the time to a 0.5 Hz deviation from more than 800 TS to about 500 TS and then 380 TS.Compromised WAMPAC propagates the attack across areas through inter-area coupling.Source: E. Evaluation of Impact Based on… · p. 7
F. Comparative Analysis: Static versus Learned ADM Boundaries
Learned ADM bounds reduce the objective and delay over-frequency threshold crossings relative to static BDD constraints, but do not prevent the attacks from reaching the threshold within the horizon. The reported comparisons span both IEEE systems.
F. Comparative Analysis: Static versus Learned ADM Boundaries
- Under static BDD, the mean generator frequency reached the 60.5 Hz over-frequency setting at 400 TS on the 39-bus system and 580 TS on the 118-bus system.The comparison holds the optimization problem and accessibility budget fixed while changing the ADM constraints.Source: F. Comparative Analysis: Static versus… · p. 7
- Compared with static BDD, learned ADM postponed 60.5 Hz crossings to 880 TS on the 39-bus and 1000–1020 TS on the 118-bus systems; the passage also reports 35.4% and 55.0%, respectively.Source: F. Comparative Analysis: Static versus… · p. 8, Figure 10 · p. 8, Figure 11 · p. 8
- Every learned-bound attack still reached 60.5 Hz within the horizon with an objective exceeding twice the benign baseline, and terminal-voltage excursions exceeded the ±10% band.The learned detector therefore delays rather than prevents the over-frequency trip.Source: F. Comparative Analysis: Static versus… · p. 8
G. Sensitivity to the Learned Detection Boundary
Learned boundaries reduce attack severity and delay over-frequency crossings but leave feasible attacks, while residual severity depends more on boundary width than detector family. The discussion also identifies cumulative voltage drift and wide-area access as security concerns, and the conclusion states the framework’s scope and future extensions.
G. Sensitivity to the Learned Detection Boundary
- On the 39-bus system, three learned detectors produce identical attacks at 60.5 Hz and 880 TS, while the wider LSTM bound raises objective by 30.8%.On the 118-bus system, the wider LSTM bound yields a larger objective (1140.90 vs. 1058.93) and earlier crossing (890 vs. 1000–1020 TS) than DBSCAN and one-class SVM.Source: G. Sensitivity to the Learned… · p. 8, Table IV · p. 9
VI. DISCUSSION
- Rate-based limits restrict rapid frequency manipulation, redirecting attacks toward excitation; because they constrain change rate rather than cumulative drift, residual voltage excursions remain similar across detectors, motivating cumulative-displacement monitoring.Learned constraints shifted solve times from 49.7 to 89.1 s on the 39-bus system and from 509.6 to 401.6 s on the 118-bus system; every solve finished under 9 min, and vulnerability and mitigation trends persisted at 118 buses.Source: G. Sensitivity to the Learned… · pp. 8–9
- Wide-area access yields 3.3× and 8.1× the benign objective on the 39- and 118-bus systems, respectively, and more than halves time to a 0.5 Hz excursion.A single wide-area signal reaches multiple generators’ excitation paths, enabling coordinated perturbation.Source: §VI · p. 9
- 15.5–55.0% objective reductions delay 60.5 Hz crossings by about 5–10 s, yet every configuration retains an optimal stealthy attack exceeding twice its benign baseline.The conclusion reports that learned boundaries reduce severity and delay crossings but do not prevent them.Source: §VI · p. 9, §VII · p. 9
VII. CONCLUSION
- The framework formulates optimal stealthy FDI as a MILP over coupled wide-area, secondary, and primary control dynamics, with the detector represented as a replaceable constraint set.Future work will add cascading failures, communication latency, and cumulative-displacement monitoring, and evaluate scalability on larger systems.Source: §VII · p. 9