Stealthy bias injection attack detection based on Kullback-Leibler divergence in noisy linear systems
summary
The gist
Stealthy bias injection attack detection based on Kullback–Leibler divergence in noisy linear systems studies the design of detection observers against stealthy bias injection attacks in stochastic…
In short
The study designs observers to detect stealthy bias injection attacks in noisy linear systems using Kullback-Leibler divergence (KLD). The goal is to maximize detection of the worst-case attacks by optimizing observer gains. The research shows that the Kalman filter is optimal at attack onset, while bi-convex approximations and ADMM/AO methods solve complex steady-state problems efficiently.
Key concepts
- Bias Injection Attack (BIA)
- This is an adversarial attack where a constant bias signal is added to a linear system. The attacker compromises some sensors and exploits the system's dynamics and noise to make the injected bias statistically indistinguishable from normal system behavior, making it very stealthy.
- Kullback-Leibler Divergence (KLD)
- KLD is a statistical measure used here to quantify how different two probability distributions are. In this context, it measures the statistical distinguishability between the residuals generated when an attack is present versus when no attack exists. A larger KLD indicates better detection performance.
- Observer Gain Design
- This refers to choosing the optimal settings for a filter (observer) that estimates the system's state based on noisy sensor measurements. The paper optimizes these gains to ensure that any subtle bias injection attack causes a significant statistical change in the observer's output, thus enhancing detection.
- Worst-Case Detection Formulation
- The optimization problem is set up to find observer gains that are best at detecting the most difficult possible attacks—those with the lowest detectability. This ensures a robust detection framework capable of identifying even highly stealthy bias injection scenarios.
Terminology used across episodes
This episode discusses
- Stealthy bias injection attack detection based on Kullback-Leibler divergence in noisy linear systems · Paper Radio
- False Data-Injection Attack Detection in Cyber-Physical Systems: A Wasserstein Distributionally Robust Reachability Optimization Approach
- Eigenvalue and Generalized Eigenvalue Problems: Tutorial
The paper
Stealthy bias injection attack detection based on Kullback-Leibler divergence in noisy linear systems · Read on arXiv
Jingwei Dong, André M. H. Teixeira
Uppsala University
Transcript
Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.
Rosa: I'm Rosa.
Dev: And I'm Dev, here with Taro.
Rosa: We're looking at the paper: "Stealthy bias injection attack detection based on Kullback-Leibler divergence in noisy linear systems".
Dev: Here's the gist of the paper. Stealthy bias injection attack detection based on Kullback–Leibler divergence in noisy linear systems studies the design of detection observers against stealthy bias injection attacks in stochastic linear systems under Gaussian noise System Model and Attack Characterization The paper considers a discrete-time LTI system driven by Gaussian white noise, where the state estimation error dynamics are modeled to indicate anomalies The bias injection attack is characterized as a step signal, where the attack vector is constant and the compromised sensors are specified by a matrix Da The impact of BIAs is quantified using a weighted norm, defined as ∥Daa¯∥2 W = ¯a⊤D⊤a W Daa¯ The stealthiness of the attack is characterized by the Kullback–Leibler divergence (KLD) between the distributions of attacked and attack-free residuals, which captures statistical distinguishability Worst-Case Detector Design Formulation The observer design problem is formulated as a max–min optimization problem to maximize the detectability of worst-case BIAs, defined as attack vectors that attain the lowest detectability while ensuring a prescribed impact on the system To obtain tractable synthesis problems, KLD-based detectability measures are considered at three specific time instants: attack onset, one step after attack occurrence, and the steady state For the attack onset case (k=0), the Kalman filter is proven to be optimal for maximizing BIA detectability regardless of the set of compromised sensors and weighting matrix Optimization Approximations for Tractability For the one-step and steady-state cases, the resulting observer design problems are inherently non-convex and are reformulated as bi-convex optimization problems Two numerical solution approaches, alternating optimization (AO) and the alternating direction method of multipliers (ADMM), are provided to solve these problems Furthermore, more tractable linear matrix inequality relaxations are developed for the one-step and steady-state cases Detection Performance Across Time Instants The paper investigates detection performance at three specific instants: attack onset, one step after attack occurrence, and the steady state The Kalman filter is optimal for the attack onset case For the one-step case (k=1), a bi-convex approximation is used, and for the steady-state case (k→∞), an LMI-based approximation is provided The KLD detector exhibits the largest KLD value as the system approaches steady state, indicating the best detection performance in that regime Simulation Validation on Thermal System The effectiveness of the proposed BIA detection methods is validated through an application to a thermal system Simulation results show that the KLD∞ detector yields the largest KLD value as the system approaches steady state, demonstrating its detection capability in the steady state The KLD1 detector attains the largest KLD value during the transient phase, showing best detection capability during this phase The KLD∞ detector exhibits the fastest KLD growth rate and the largest steady-state KLD value, indicating the best detection performance The paper concludes that these proposed detectors are able to detect worst-case attacks, while traditional Kalman filter fails to do so The paper proposes a systematic and theoretically grounded approach for stealth-aware detector design in stochastic linear systems Future work will focus on extending the proposed framework to incorporate active detection mechanisms to address broader classes of adversarial attacks The results offer a flexible way to balance transient and steady-state detectability by solving a multi-objective optimization problem The derived bi-convex optimization problem (21) can be efficiently solved using AO and ADMM algorithms The LMI-based optimization problem serves as an effective initialization for the AO and ADMM algorithms Appendix A Technical Proofs The appendix provides useful matrix relations and inequalities instrumental in later proofs Lemma A.1 provides Young’s relation, which is used to establish the sufficiency of the conditions in (18) for (14) The proof of Theorem 3.6 applies Schur complement to derive the required inequality, proving the sufficiency of the derived conditions in (18) for (14) The proof of Lemma 3.7 provides an alternative expression for the steady-state transition matrix Φ(∞, L), which is crucial for deriving the bi-convex approximation The proof of Theorem 3.9 applies Schur complement to derive the final inequality, demonstrating that constraints in (19) are satisfied under conditions in (23) The proof of Theorem 3.8 utilizes Young’s relation and Schur complement to show that the constraints in (19) are satisfied The proof of Lemma 3.2 establishes the lower bound on the inverse covariance matrix, which is essential for solving the optimization problem The proof of Proposition 3.4 shows that the Kalman filter is optimal for onset detectability regardless of Da and W The proof of Theorem 3.5 demonstrates that the bi-convex approximation provides a lower bound on the objective value, which is sufficient to solve the problem The proof of Theorem 3.8 confirms that the LMI-based approximation provides a feasible solution for steady-state detectability The proof of Theorem 3.9 demonstrates that the constraints in (19) are satisfied under conditions in (23) The proof of Lemma 3.3 establishes the monotonicity property of the smallest generalized eigenvalues, which is instrumental in establishing the optimality of the Kalman filter at attack onset The proof regarding the transformation of Φ(∞, L) yields a simplified form for steady-state analysis The proof regarding Schur complement in Theorem 3.9 shows that the constraints in (19) are satisfied under conditions in (23) The proof of Lemma A.1 provides matrix relations used to derive the final inequality for the LMI-based approximation The proof of Theorem 3.6 shows that the LMI-based approximation can be solved efficiently using AO and ADMM The proof of Algorithm 2 outlines the steps for solving the optimization problems using ADMM The proof regarding the detection logic shows that I(k) ≤ τ, normal, I(k) > τ, attack detected The simulation results show that the Mahalanobis distance exceeds the threshold after attack onset and remains below it for optimal attacks The KLD∞ detector exhibits the fastest KLD growth rate and the largest steady-state KLD value, indicating the best detection performance The simulation results show that the Mahalanobis distance generated by traditional Kalman filter oscillates around the threshold after attack occurs, making it unable to detect the attack timely References
1: S.
Rosa: H.
Dev: Ahmed, G.
Rosa: Kim, D.
Paper summary: Rosa: So, wrapping up the paper "Stealthy bias injection attack detection based on Kullback–Leibler divergence in noisy linear systems," what we've seen is their systematic approach to designing these stealth-aware observers for stochastic linear systems >
Dev: They focused on using KLD to measure statistical distinguishability between attacked and unattacked states, which they then optimized using a max-min framework against worst-case attacks >
Taro: The paper tackles the practical challenge of making this optimization tractable by breaking it down into analysis at attack onset, one step after the attack, and steady state >
Rosa: They show that the Kalman filter is optimal for detecting those attacks right when they start, which is a strong starting point for any detection strategy >
Dev: For the later stages of the attack, they use bi-convex approximations solved with alternating optimization and ADMM methods to get solutions, plus some linear matrix inequality relaxations >
Taro: The implication is that we can get a detector that balances transient detection—catching the attack early—with steady-state detection—seeing what's happening long term >
Rosa: It means for someone listening who only cares about the big picture, this work suggests a way to design something that doesn't just look for obvious errors but looks at how the entire system's statistics are shifting due to a hidden attack >
Conclusion: Rosa: So, we've looked at how they design these observers to spot stealthy bias injection attacks using Kullback–Leibler divergence in noisy linear systems >
Dev: Yeah, it’s about making sure the detection works even when an attacker is trying to be really quiet by just messing with a few sensors and the system dynamics >
Taro: It gets interesting because they aren't just looking for big errors, they're measuring how statistically different the noise looks before and after the attack >
Rosa: The authors, Ahmed, Kim, Kim and Peng Sun et al., they laid out a whole framework for this detection using that divergence measure >
Dev: It’s not some abstract theory; it’s about designing an observer gain setup where you maximize detectability against the worst possible attacks given a certain impact constraint >
Taro: They specifically look at three points in time for when to check the KLD—right when the attack starts, one step later, and then when things settle down >
Rosa: And what they found is that for detecting an attack right at onset, the standard Kalman filter is actually optimal no matter which sensors are compromised >
Dev: That's a big deal because it means we don't need super complex stuff just to catch the initial intrusion, even if the system dynamics are weird >
Taro: But then for steady state detection, things get trickier and they use bi-convex approximations solved with things like alternating optimization and ADMM methods >
Rosa: So what this means for us outside the lab is that we can design systems that are sensitive to these stealthy intrusions in real-world conditions >
Dev: It changes how we think about safety cases; instead of just checking if the system state is within bounds, you're checking the statistical signature of the noise itself >
Taro: We’ve seen simulations on a thermal system where they show that using their steady-state detector gives you the biggest KLD value when things are stable >
Rosa: It shows that even in a quiet, long-running system, if you use this method, it can still pick up subtle shifts caused by those kind of attacks >
Dev: But we still need to figure out how to make those optimization problems run fast enough for real-time control loops >
Taro: That's what we’ll be looking at next; the practical implementation and how it handles the actual failure modes when things go sideways >
More episodes
- 2608.13438-ContactGuard: Pre-Contact Execution Monitoring with Action-Conditioned Latent World Models
- 2610.34593-Future Prediction as Training-Only Video Supervision for Pretrained VLA Policies
- 2610.35385-EmbodiedLab: Beyond Manipulation, Benchmarking Embodied Agents That Act to Know
- 2610.36378-Data-Driven Hamiltonian for Neural Optimal Control
- 2610.37275-A Heuristic for Scheduling Autonomous Vehicles at Intersections
- 2610.37288-Dynamic Coordination of Autonomous Vehicles at Intersections
- 2610.37632-Real-Time Coordination of Autonomous Vehicles at Intersections
- 2611.13779-Softness is something you also need! Why compliance helps control: information metrics of a passive walker with compliant ankle
- 2611.13411-Residual-Corrected Physics-Informed Neural Network KKL Observers with Input-to-State Stability Guarantees
- 2611.14488-Is Your Robot's Memory Just a Clock? Temporal Shortcuts in Memory-Augmented VLA Policies