Source-linked AI summary
DynaTrust-VVC: Directional Physics-Informed Trust-Based Detection and Mitigation for Cyber-Resilient Multi-Agent Volt--VAR Control
Md Fazley Rafy, Kamrul Hasan, Anurag K. Srivastava
TL;DR
Distributed Volt–VAR control relies on voltage measurements that false-data injection can corrupt, and coordinated neighbors can falsely corroborate one another. DynaTrust-VVC combines directional physics-informed trust with adaptive evidence gating and counterfactual safe-voltage recovery. On the nonlinear IEEE 123-bus feeder, it detects the coordinated three-bus attack one minute after onset in all 27 matched replicates, while the composite resilience index rises from 0.227 to 0.315 without a statistically significant increase in benign false-alarm rate.
Problem
Distributed Volt–VAR control depends on voltage measurements vulnerable to false-data injection, and coordinated compromised neighbors can falsely corroborate one another.
Source: §I · p. 1Method
DynaTrust-VVC assigns incoming neighbor messages directional physics-informed trust and combines it with adaptive evidence gating and counterfactual safe-voltage recovery.
Source: §I · p. 2, §II · p. 3, A. Proposed DynaTrust-VVC Extensions · p. 3Results
DynaTrust-VVC detects the coordinated three-bus attack one minute after onset in all 27 matched replicates, and the composite resilience index rises from 0.227 to 0.315.
Source: §IV · p. 5, §V · p. 6Takeaways & Limitations
The framework lets agents assess whether neighboring messages are physically consistent before using them in control decisions.
Source: §V · p. 6
Abstract
from arXiv · showhide
Distributed Volt--VAR control relies on voltage measurements and is therefore vulnerable to false-data injection. Neighbor corroboration can distinguish an isolated corrupted measurement from a physical disturbance, but coordinated agents can falsely corroborate one another. This paper proposes DynaTrust-VVC, a cyber-resilient multi-agent Volt--VAR framework that assigns each incoming neighbor message a directional trust value based on the consistency between its reported voltage increment and a sensitivity-based prediction from active- and reactive-power changes. Trust falls immediately after a physics mismatch and recovers gradually after sustained consistency. DynaTrust-VVC further uses a physics-adaptive evidence threshold and counterfactual safe-voltage recovery-based mitigation. The proposed method was validated using a nonlinear IEEE 123-bus feeder with minute-resolution residential profiles and 27 matched replicates. DynaTrust-VVC identifies a coordinated three-bus attack one minute after onset in all 27 matched replicates, whereas an otherwise identical configuration with fixed neighbor trust identifies none. The composite resilience index also increases from 0.227 to 0.315, with no statistically significant increase in the benign false-alarm rate.
I. INTRODUCTION
Distributed Volt–VAR control depends on voltage measurements that false-data injection can corrupt, while coordinated compromised neighbors can defeat ordinary corroboration. DynaTrust-VVC addresses this with directional physics-based trust, adaptive evidence gating, and counterfactual safe-voltage recovery.
I. INTRODUCTION
- Coordinated compromised neighbors can corroborate falsified measurements, while centralized detection adds latency and single points of failure and many existing methods operate outside the control loop.Learning-based multi-agent VVC methods primarily optimize regulation without explicitly addressing attack detection, localization, or mitigation.Source: §I · p. 1
- DynaTrust-VVC replaces static corroboration weights with per-link directional trust based on whether reported voltage changes match sensitivity-based predictions from active- and reactive-power changes.Trust drops immediately after a physics mismatch and recovers gradually after sustained consistency.Source: §I · pp. 1–2
- The framework combines physics-adaptive evidence gating with counterfactual safe-voltage recovery in the distributed Volt–VAR control loop.An agent estimates voltage without the corrupted measurement and adapts its reactive-power command using trusted power changes.Source: §I · p. 2
- The evaluation uses 27 matched replicates, internal module ablation, and a sensitivity-mismatch study to quantify component contributions and boundaries.The introduction also identifies distinguishing malicious manipulation from legitimate physical disturbances while remaining lightweight enough for inverter control loops as a defense requirement.Source: §I · pp. 1–2
II. PROPOSED METHODOLOGY
The methodology combines local anomaly evidence and neighbor corroboration within a distributed Volt–VAR control layer. DynaTrust-VVC extends that layer with directional physics-informed trust, adaptive evidence gating, and counterfactual safe-voltage recovery.
II. PROPOSED METHODOLOGY
- The threat model alters measured voltage through bias, ramp, replay, intermittent, or noise attacks while leaving feeder physics and actual power injections unchanged.Agents share scalar phase-aggregated quantities with electrical neighbors.Source: §II · p. 2
- Local evidence combines autoencoder reconstruction error with discrepancy from the last trusted safe-state voltage, and persistence requires two consecutive minutes above the trip level with a consistent discrepancy sign.The safe-state voltage remains frozen while cyber suspicion is latched.Source: §II · p. 2
- Neighbor compatibility is high when reported discrepancy and normalized reactive-power changes agree; corroboration supports a physical disturbance, whereas high local evidence with low corroboration supports localized corruption.A sign-compatibility factor is 1 for matching nonzero signs or when either discrepancy is zero, and 0.35 otherwise.Source: §II · p. 2
- A suspicion latch activates after persistent local evidence exceeds its trip level and releases when corroboration exceeds the physical-release level or the latch horizon expires.The diagnostic score indicates support for a cyber event, while a separate gain controls movement of the reactive-power command toward its safe target.Source: §II · p. 2
- The final target blends the nominal command with a safe command calculated from a trusted safe-voltage estimate, or uses the stored pre-suspicion command when that estimate is unavailable.The final command is rate-limited by a fraction of available reactive capability and projected onto that capability interval.Source: §II · pp. 2–3
- DynaTrust-VVC extends the common edge anomaly-detection, neighbor-evidence, and Volt–VAR control layer with directional physics-informed trust, adaptive evidence gating, and counterfactual safe-voltage recovery.DynaTrust names the time-varying trust assigned to each incoming neighbor message.Source: §II · p. 3
A. Proposed DynaTrust-VVC Extensions
DynaTrust-VVC scores each neighbor message against a sensitivity-based voltage-increment prediction, then combines directional trust with adaptive evidence gating and counterfactual safe-voltage recovery.
Directional physics-informed trust
- Each agent predicts a neighbor’s reported voltage increment using a first-order sensitivity model and evaluates the normalized physics residual against an allowable scale.The model uses active- and reactive-power voltage sensitivities.Source: A. Proposed DynaTrust-VVC Extensions · p. 3
- Trust drops immediately when a message’s instantaneous trust falls below its stored value and recovers gradually toward one after consistency.Trust states start at one; λ_τ controls recovery.Source: A. Proposed DynaTrust-VVC Extensions · pp. 3–4
Physics-adaptive evidence gating
- The expected voltage excursion adapts the evidence-trip level, but this extension changes the evidence gate only, not the autoencoder threshold.The trip level is bounded by ζ_max.Source: A. Proposed DynaTrust-VVC Extensions · p. 4
Counterfactual safe-voltage recovery
- When the cyber latch is active, the method estimates a counterfactual safe voltage and evaluates the droop law at that voltage for its safe command.The estimate is initialized at latch onset and reset when the latch is released; insufficient trusted-peer support leads to using the last trusted command.Source: A. Proposed DynaTrust-VVC Extensions · p. 4
III. EXPERIMENTAL DESIGN
The evaluation uses minute-resolution simulations of a nonlinear IEEE 123-bus feeder with residential load and PV profiles, divided into chronological training, validation, and test sets.
Feeder and data
- The nonlinear IEEE 123-bus feeder was solved at one-minute resolution with ten controllable PV inverters and residential load and PV profiles.Of 281 complete common days, 168 were used for training, 56 for validation, and 57 for testing.Source: §III · p. 4
Parameter provenance
- Table I distinguishes parameters selected on validation days, derived from physics, and fixed as design constants.The table labels these categories V, P, and D, respectively.Source: Table I · p. 4
Evaluation scenarios
- The primary evaluation scenarios use an attack window from minute 360 through minute 600.Table II lists the primary scenarios.Source: Table II · p. 4, §III · p. 4
Feeder and data
- All defenses used ten homogeneous agents connected by directed three-neighbor links.The arrangement had previously been demonstrated using distributed embedded controllers.Source: §III · p. 4
A. Configurations, Metrics, and Statistical Protocol
The evaluation compares static-trust and progressively extended configurations, including the full DynaTrust-VVC method, and specifies a sensitivity-mismatch variant.
Configurations
- Static-trust SA disables the three DynaTrust extensions by fixing neighbor trust at one, retaining the nominal evidence trip, and using the held command as the safe command.The table compares SA with the proposed DT configuration over 27 matched replicates.Source: A. Configurations, Metrics, and… · p. 4, Table III · p. 5
- The module chain adds physics trust to SA, then adaptive thresholding, and finally counterfactual recovery in full DT.The sensitivity-mismatch variant perturbs each agent-side sensitivity entry by a factor drawn from U[0.8, 1.2].Source: A. Configurations, Metrics, and… · p. 4
Per-agent procedure
- Each agent generates local evidence, broadcasts it, updates neighbor trust, computes adaptive gating and recovery, then rate-limits and projects its reactive-power target.The sequence uses the per-agent steps summarized in Algorithm 1.Source: A. Configurations, Metrics, and… · p. 4
D Mraw CRVI
The evaluation scores detection, localization, mitigation, availability, benign false alarms, and composite resilience, using matched replicates and paired bootstrap intervals for principal comparisons.
Metrics
- Detection score D uses first-alarm latency measured from attack onset, with the decay constant fixed at 15 minutes before test evaluation.Source: D Mraw CRVI · p. 5
- Localization L is the balanced average of recall and specificity.Source: D Mraw CRVI · p. 5
- Mitigation uses a bounded score based on weighted voltage–reactive-power tracking cost relative to the clean run, alongside an unbounded raw measure that exposes degradation hidden by clipping.Availability is the mean fraction of feeder buses within [0.95, 1.05] p.u. during the attack window; benign FAR is the fraction of node-minutes meeting the cyber-evidence threshold.Source: D Mraw CRVI · p. 5
Statistical protocol
- The composite resilience index is a weighted geometric mean, and comparisons use 27 matched replicates with paired-bootstrap 95% confidence intervals for principal comparisons.The replicates combine nine random seeds across three held-out test days; intervals use 20,000 resamples.Source: D Mraw CRVI · p. 5
IV. RESULT ANALYSIS
DynaTrust-VVC detects coordinated attacks that static trust misses, while its directional trust and counterfactual recovery improve resilience and mitigation. The results also expose limits for ramp and replay attacks.
IV. RESULT ANALYSIS
- For the coordinated-corridor attack, SA fails in all 27 replicates, while DT detects it one minute after onset and raises scenario CRVI from 0.007 to 0.290.Source: §IV · p. 5
- Directional trust raises pooled D from 0.234 for SA to 0.468 after PT; full DT raises corridor raw mitigation from +0.003 for PT+AT to +0.028.Source: Table IV · p. 5, §IV · p. 6
- Across four benchmark attacks, CRVI rises from 0.227 for SA to 0.315 for DT; paired change is +0.101 (95% CI [0.079, 0.127]) in all 27 replicates.Source: §IV · p. 5
- In a seed-0 example, a shared −0.035 p.u. step without P–Q change retains unit trust in SA, while DT’s residual triggers detection one minute after onset.Source: §IV · p. 5
- With ±20% sensitivity perturbation, coordinated detection remains 0.936 and raw mitigation +0.028, while benign FAR rises from 0.107 to 0.125.Source: §IV · p. 6
- Ramp and replay remain difficult for SA and DT because their signatures are insufficiently distinct from nominal load–PV variation over the 240-minute attack window.Source: §IV · p. 6
V. CONCLUSIONS
DynaTrust-VVC combines directional physics-informed trust, adaptive evidence gating, and counterfactual recovery in multi-agent Volt–VAR control. On the IEEE 123-bus feeder, it detects coordinated attacks promptly and improves composite resilience, though evaluation and sensitivity estimates remain bounded.
V. CONCLUSIONS
- DynaTrust-VVC detects the coordinated three-bus attack one minute after onset in all 27 matched replicates, while composite resilience increases from 0.227 to 0.315 without a statistically significant benign false-alarm increase.Source: §V · p. 6
- The framework combines directional physics-informed trust, adaptive evidence gating, and counterfactual safe-voltage recovery so agents assess neighboring messages for physical consistency before control decisions.Source: §V · p. 6
- Evaluation is limited to the IEEE 123-bus feeder, and deployment requires sensitivity estimates reliable as feeder topology and operating conditions change.Source: §V · p. 6