Secure Filtering against Spatio-Temporal False Data Attacks under Asynchronous Sampling

1. Problem Statement

The paper addresses the secure state estimation problem for continuous linear time-invariant (LTI) systems equipped with sensors that operate under non-periodic and asynchronous sampling. In this setup, sensors transmit measurement triples consisting of the sensor index, the sampling time-stamp, and the measurement value to a fusion center.

The system is vulnerable to spatio-temporal false data attacks, where an adversary can manipulate the measurement triples in four distinct ways (or a combination thereof):

1. False-data injection: Altering the measurement value while keeping the correct time-stamp.
2. Time-stamp manipulation: Altering the time-stamp while keeping the measurement value correct (causing temporal misalignment).
3. Denial-of-Service (DoS): Blocking the transmission of measurement triples.
4. False-data generation: Injecting completely new, fabricated measurement triples with arbitrary time-stamps.

The goal is to design a decentralized estimation algorithm that provides a state estimate with a uniformly bounded error, regardless of the presence of these attacks, provided the number of compromised sensors is below a certain threshold ($p$-sparse attacks).

2. Methodology

The proposed solution is a decentralized estimation algorithm based on the decomposition of the Kalman Filter (KF) and a robust optimization fusion scheme.

A. Asynchronous Sampled-Data KF Decomposition

The authors first establish a baseline "Oracle KF" (assuming no attacks) for asynchronous systems. They decompose this global KF into a set of local estimators ($\zeta_i[k]$) running at each sensor.

• The global state estimate is shown to be the sum of these local estimates: $\hat{x}[k] = \sum \zeta_i[k]$.
• The local estimates are stable and converge to a linear transformation of the true state ($G_i[k]x[k]$).
• This decomposition allows the system to isolate the impact of attacks to specific local estimators corresponding to compromised sensors.

B. Weighted Least Squares (WLS) Fusion

In the absence of attacks, the authors propose fusing the local estimates using a Weighted Least Squares (WLS) optimization problem.

• Objective: Minimize a weighted error term subject to the constraint that the fused state matches the local estimates.
• Result: Theorem 1 proves that the solution to this WLS problem exactly recovers the optimal state estimate of the global Oracle KF.

C. Secure Fusion with $\ell_1$ Regularization

To mitigate attacks, the WLS problem is modified by introducing an $\ell_1$ regularization term (similar to Lasso regression).

• Optimization Problem:
  $$\min_{\check{x}, \mu, \vartheta} \frac{1}{2}\mu^\top \tilde{W}^{-1} \mu + \gamma \|\vartheta\|_1$$
  Subject to: $V[k]\zeta[k] = H\check{x}[k] + \mu[k] + \vartheta[k]$
  • $\check{x}$: The secure state estimate.
  • $\mu$: Represents the estimation error due to process/measurement noise.
  • $\vartheta$: Represents the sparse attack vector. The $\ell_1$ norm encourages sparsity, effectively forcing the algorithm to identify and isolate the corrupted sensors (where $\vartheta \neq 0$) while ignoring their malicious data.
• Mechanism: The regularization parameter $\gamma$ is tuned such that if the number of attacked sensors is within the observability redundancy limit, the algorithm correctly identifies the attack vector and excludes it from the state calculation.

3. Key Contributions

1. Novel Attack Model: The paper unifies integrity attacks (injection), availability attacks (DoS), and temporal attacks (time-stamp manipulation) into a single spatio-temporal false data attack framework. It specifically addresses the challenge of time-stamp manipulation in asynchronous systems, a gap in existing literature.
2. Decentralized Secure Estimator: The authors propose a novel algorithm that:
   • Recovers the optimal Kalman estimate exactly when no attacks are present.
   • Provides secure estimates with uniformly bounded errors under attacks, where the error bound is independent of the attack magnitude and depends only on system parameters and the "oracle" performance.
3. Theoretical Guarantees:
   • Theorem 2: Proves that under a specific condition on $\gamma$, the secure estimator is identical to the optimal KF in the absence of attacks.
   • Theorem 3: Proves that in the presence of $p$-sparse attacks, the estimation error is bounded. The bound relies on the system being $2p$-sparse observable (a standard redundancy assumption in secure estimation).
4. Handling Time-Stamp Manipulation: The paper demonstrates that time-stamp manipulation can be mathematically modeled as a measurement mismatch (effectively a false-data injection), allowing the $\ell_1$-regularized fusion to handle it alongside other attack types.

4. Results

The effectiveness of the algorithm was validated through simulations on the IEEE 14-bus power system (28 state variables, 42 sensors).

• Scenario 1 (No Attacks): The proposed secure estimator (with appropriate $\gamma$) produced results identical to the standard Kalman Filter and the baseline WLS, confirming optimality in attack-free conditions.
• Scenario 2 (Under Attack): The system was subjected to a mix of attacks:
  • False-data injection on Bus 3.
  • Time-stamp manipulation on Bus 5.
  • DoS on Bus 4.
  • False-data generation on Bus 2.
• Performance:
  • The standard Kalman Filter failed catastrophically, showing large estimation errors.
  • A Robust Kalman Filter (existing method) performed better but still exhibited significant errors.
  • The proposed secure estimator maintained a state estimate close to the true state, with errors remaining within the theoretical upper bound derived in Theorem 3. It successfully isolated the corrupted sensors and mitigated the impact of all four attack types simultaneously.

5. Significance

This paper makes a significant contribution to the field of Cyber-Physical Systems (CPS) security by addressing a realistic and complex threat model.

• Realism: Unlike many prior works that assume synchronous, periodic sampling, this work tackles asynchronous and non-periodic data, which is common in large-scale distributed systems (e.g., power grids, water networks) but introduces unique vulnerabilities like time-stamp spoofing.
• Robustness: By combining decentralized estimation with $\ell_1$ regularization, the method offers a rigorous mathematical guarantee of security without requiring a central trusted authority to verify every sensor.
• Optimality Preservation: A key achievement is that the security mechanism does not degrade performance when the system is safe; it reverts to the optimal Kalman Filter, avoiding the "conservatism" often found in robust estimation methods (like median-based filters).

In summary, the paper provides a theoretically sound and practically validated framework for securing state estimation in modern, asynchronous control systems against sophisticated, multi-vector cyber-attacks.