Research
zkSTAR: A zero knowledge system for time series attack detection enforcing regulatory compliance in critical infrastructure networks
Authors are affiliated with the School of Industrial Engineering and Management, Oklahoma State University, Stillwater, Oklahoma, USA. The paper is arXiv:2510.23060v4 [cs.CR].

- arXiv
- 2510.23060
- Published
- 2025-10-27
- Authors
- Paritosh Ramanan, Sathwik Yamana, H. M. Mohaimanul Islam, Abhiram Reddy Alugula
AI summary
zkSTAR: A Zero Knowledge System for Time Series Attack Detection Enforcing Regulatory Compliance in Critical Infrastructure NetworksOverview
Research area: Applied cryptography and industrial control system (ICS) security — specifically the use of zk-SNARKs to make statistical attack-detection outcomes publicly verifiable for regulators without disclosing utility operational data.
Technical level: Advanced. The paper assumes familiarity with extended Kalman filters, residual-based χ² hypothesis testing, hash commitments, and zk-SNARK proving systems (Setup/Prove/Verify, knowledge-soundness, common reference strings).
Scope (one sentence): The paper designs and formally analyzes zkSTAR, a two-part zero-knowledge architecture that proves temporal and statistical consistency of EKF-based χ² intrusion detection for critical infrastructure networks, and positions it as a compliance-enforcement layer between utilities and regulators such as ISACs.
Authors are affiliated with the School of Industrial Engineering and Management, Oklahoma State University, Stillwater, Oklahoma, USA. The paper is arXiv:2510.23060v4 [cs.CR].
What This Paper Is About
Critical infrastructure operators (power grids, water systems, gas pipelines) detect cyberattacks locally using state-space models and statistical hypothesis tests, then report alarms upward to regulators and information-sharing bodies. Regulators currently have no way to check that those alarms are genuine without demanding raw operational data, which utilities are unwilling to hand over. zkSTAR's goal is to let a utility prove, in zero knowledge, that its detection pipeline ran correctly — that state estimates were computed recursively and that alarms followed a valid χ² test — so that a compromised utility cannot falsify or suppress alarms.
Key Contributions
-
A zk-SNARK prediction framework built on an extended Kalman filter. The authors encapsulate a pre-trained, non-linear EKF state-space model in a zero-knowledge circuit so that data-driven attack detections on utility ICS become provable while the local state-space dynamics remain hidden.
-
Formal security guarantees for regulatory compliance. They define temporal consistency (TC) and statistical consistency (SC), prove conditions under which each holds (Theorems IV.1 and IV.2), and prove a detection-suppression bound (Theorem IV.3) showing that an under-attack utility must either fail proof verification or violate the TC–SC relation.
-
A kernel decomposition scheme for scalability. Because circuit size grows exponentially with detection-window size, they split EKF inference into separate zkSNARK kernels — a Kalman Reconstruction Check (KRC) kernel and an EKF kernel — operating over batched residual outputs, plus an offline/online split for operations that are prohibitively expensive in-circuit.
-
A temporal decomposition scheme and containerized REST-API testbed. Detection horizons are divided into detection windows and finer TC intervals; the framework is exercised through case studies on real-world ICS datasets using a REST-API-driven containerized testbed.
Main Findings
-
Temporal consistency has a clean proof condition. Theorem IV.1 states that temporal consistency holds across every consecutive timestep pair t, t+1, given sensor measurements y_{t+1}, if and only if the hash commitment of the EKF input at t+1 equals the commitment of the EKF output at t and the proof at t+1 verifies (Φ_{t+1} = 1). If temporally inconsistent values are fed into EKF predictions, verification fails for at least one timestep.
-
Statistical consistency follows from temporal consistency, with a quantified probability. Theorem IV.2 states that temporal consistency of the system operating on y^rep_{t+1} is necessary and sufficient for statistical consistency with respect to y^real_{t+1} to hold with probability at least 1 − ε_max − α. Statistical consistency itself is defined as ρ^real_t = ρ^rep_t, i.e. the reported alarm bit matches the true alarm bit.
-
Alarm suppression is bounded. Theorem IV.3 gives Pr[Ψ^rep_t = 1 | Φ^rep_t = 1] ≤ ε_max + α + negl(λ) for any timestep after the attack begins, for a probabilistic polynomial-time adversary against a knowledge-sound proving system (Halo2 backend). Reusing stale states or forging alarms therefore contradicts knowledge-soundness.
-
Approximate agreement is deliberately permitted. The paper does not require bit-exact reproduction of sensor values. "Probabilistic similarity" (Definition 1) treats two measurement vectors as similar when both test statistics exceed χ²_{p,α} or both fall at or below it — acknowledging calibration error and finite-precision effects on inexpensive ICS hardware.
-
Assumption-based security envelope. Three assumptions frame the guarantees: additive perturbation of the reported measurement (y^rep_t = y^real_t + Δ_y), an operational-significance threshold ‖r^real_t‖ ≥ Γ with Γ > 0, and an EKF sensitivity bound ε_max on the probability that a reported measurement yields an outcome inconsistent with the true post-attack measurement, with ε_max ≪ α for a well-formed detector. The paper explicitly defers formal bounding of ε_max in terms of attack magnitude and EKF covariance structure to future work.
-
Empirical results are not reported in the supplied content. The abstract states that practical feasibility is validated through computational experiments on real-world ICS datasets and that case studies were run on a REST-API-driven containerized testbed, but no dataset names, dataset sizes, benchmark names, proof-generation times, or verification times appear in the text available for this summary. The paper body is truncated during the description of the EKF kernel (Algorithm 2).
-
Setup is per-utility, not per-timestep. The zk-SNARK Setup phase (1^λ, M, θ) → (pk, vk) is assumed to run once per utility, since a pre-trained high-accuracy state-space model does not need frequent weight updates; it must be re-run whenever θ changes.
Methodology in Plain English
The starting point is the standard way ICS operators detect attacks: a model predicts what sensors should read, the difference between prediction and reality is the "residual," and a statistical test (a χ² test on residuals, computed after a PCA-style decomposition of the residual covariance) raises an alarm when the residual is too large.
The problem is that to check such an alarm, a regulator would need the residuals and covariance matrices — exactly the data utilities refuse to share. So the authors wrap the whole computation in a zero-knowledge proof system. The utility commits in advance to its trained model. Then, on demand, a regulator asks for a proof covering a specific detection window, and the utility produces a proof that the pipeline executed correctly.
Two things must be proven. First, temporal consistency: the state estimate fed into timestep t+1 is genuinely the output of timestep t, which is enforced by chaining hash commitments (with nonces, so the regulator cannot brute-force the small, bounded state values) and by requiring each proof to verify. Second, statistical consistency: the alarm bit reported matches what the χ² test would actually have produced.
To make this tractable, the authors break the problem into pieces. A detection window is divided into W temporal-consistency intervals, each of D consecutive timestamps; residuals and covariances are summed across these intervals before the test is applied. Expensive linear algebra — matrix inversions, SVD decompositions, derivative evaluations — is computed outside the circuit and then checked inside it: the KRC kernel verifies the Kalman-gain relation ‖KS − PHᵀ‖²₂ < ε, and the EKF kernel chains the prediction, residual, covariance, and update equations together with that check. Proofs are generated lazily, only when a regulator requests them, which keeps utility-side overhead low.
Why This Matters
Impact on research. The paper argues that no existing zk-SNARK framework targets verification of dynamical-system integrity or hypothesis-test correctness — prior verifiable-ML work covers training integrity and model ownership, and verifiable-differential-privacy and secure-multiparty work covers confidentiality but not detection correctness. zkSTAR is presented as filling that gap, connecting the formal-methods/cryptography community with control-systems security.
Real-world applications:
- Power grid SCADA monitoring — proving to a regulator that false-data-injection alarms on substation sensors were genuine and not suppressed during an incident.
- Water and wastewater treatment control — providing auditable attack-detection evidence without exposing pump, valve, and chemical-dosing telemetry.
- Gas pipeline telemetry — corroborating alarms during incidents where adversaries have simultaneously compromised IT and OT layers.
- ISAC network-wide situational awareness — letting information-sharing centers aggregate detection reports whose integrity is cryptographically attested rather than merely asserted, reducing false positives and unnecessary service disruptions.
The threat motivation is concrete: the paper cites Industroyer and Industroyer2, Triton, and Stuxnet as demonstrations that adversaries can manipulate control systems and monitoring data jointly.
Industry relevance. The framework targets the compliance bottleneck directly — replacing costly, slow, privacy-intrusive audits with on-demand cryptographic attestation — while the lazy proof-generation design is aimed at keeping the burden on utilities low. The authors assume an honest-but-curious verifier and note that channels need integrity, achievable via TLS, rather than confidentiality.
Future Directions
-
Formally bound ε_max. The authors explicitly defer deriving ε_max in terms of attack magnitude and EKF covariance structure to future work; this is the key missing piece linking the cryptographic guarantee to physical damage thresholds.
-
Extend beyond false-data-injection attacks. The threat model excludes gradual-degradation attacks that require persistent asset control over weekly or monthly horizons, and assumes adversaries lack long-term control of sensing and control hardware. Covering those cases is an open problem.
-
Address the circuit-size bottleneck. The paper acknowledges that zkSNARK circuit size grows exponentially with detection-window size, which is why the kernel decomposition and the compute-offline–validate-online split exist. Whether this scales to large, fine-granularity streaming datasets across many utilities is not settled in the available text.
-
Tooling gaps for in-circuit numerics. The authors note that current zkSNARK frameworks such as ezkl lack support for linear solvers, SVD decompositions, and derivative evaluations of complex sub-modules due to their cost in zero knowledge; better native support would remove the need for off-circuit computation and its associated trust assumptions.
Target Audience
Researchers and graduate students working at the intersection of applied cryptography and industrial control systems security; security architects at utilities and transmission system operators evaluating privacy-preserving compliance mechanisms; regulators and ISAC staff interested in what cryptographic attestation can and cannot prove about a utility's detection pipeline; and zk-SNARK practitioners looking for a non-financial application domain with hard real-time and numeric-precision constraints. Readers without background in Kalman filtering or proof systems will find the formal sections (III and IV) demanding.
Authors’ abstract
Industrial control systems (ICS) form the operational backbone of critical infrastructure networks (CIN) such as power grids, water supply systems, and gas pipelines. As cyber threats to these systems escalate, regulatory agencies are imposing stricter compliance requirements to ensure system-wide security and reliability. A central challenge, however, is enabling regulators to verify the effectiveness of detection mechanisms without requiring utilities to disclose sensitive…
The full abstract on arXiv