Pith. sign in

REVIEW 4 major objections 4 minor 26 references

Online Detection and Mitigation of Robust Zero Dynamics Anomaly Behavior in MIMO Nonlinear Control Systems

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims that robust zero dynamics attacks in MIMO nonlinear systems can be detected from the difference between the plant input and a nominal model's controller output, and then neutralized by an online-adapted neural network…

desk verdict Flawed proofs undercut an otherwise timely attack detection/recovery scheme. read the letter →

arxiv 2506.03298 v1 pith:SRI66C2K submitted 2025-06-03 eess.SY cs.SY

Kosar Behnia , H.A. Talebi , Farzaneh Abdollahi This is my paper
classification eess.SYcs.SY
keywords cyber-physicalsystemsrobustzerodynamicsattacksattackdetectionmitigationresidualsignalneuralnetworknonlinearMIMOLyapunovstability
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tackles robust zero dynamics attacks, a class of stealthy cyber attacks that drive a system's internal (zero) dynamics unstable while leaving its measured outputs nearly unchanged, making them hard to detect. The authors propose a two-stage strategy that works without an exact plant model: a residual signal built from the difference between the plant input and the output of a nominal model's controller flags the attack, then a neural network, trained online with no prior data, estimates the attack signal and subtracts it from the input to recover the closed-loop system. The core claim is that the detection residual converges to zero in the absence of attack and to the attack signal under attack, and that the recovery loop keeps the estimation error bounded, the zero dynamics bounded, and the output tracking the reference. If correct, this would extend zero-dynamics attack defense from linear or exactly modeled systems to general MIMO nonlinear systems with model uncertainty and time-varying references.

What carries the argument

The carrying object is the residual signal $r := u_c + \alpha - u_{mc}$, formed by feeding the real system output into a nominal model with its own controller and subtracting that controller's output from the actual plant input; this single signal does double duty, serving as detection statistic and as the online training error for the recovery network. The recovery network is a one-hidden-layer neural network whose hidden-layer activation is the nominal model's $F_n$ and whose output layer is the known map $G_n^{-1}$, so its output is $z_r = G_n^{-1}(F_n(w,x^*))$ and its weights $w$ evolve by the adaptation rule in (18), driven by $e = \alpha - G_n^{-1}(F_n(w,x^*)) + u_c - u_{mc}$. The proofs are carried by a Lyapunov function $V' = V_\sigma(x,z',s) + (\alpha-z_r)^T(\alpha-z_r) + L^T L$ with an auxiliary variable $L$ defined in (29), which is used to absorb the cross-terms between the estimation error and the residual.

What would settle it

One concrete check: in the attack-free four-tank setting of Section V, compute the right-hand side of (17) along the simulated error trajectory. If $\dot{V}_\sigma$ is not negative despite $c_3 > 6c_5c_4$ -- because the positive term $6c_5c_4 \|[z,x,s]\|^2$ involving the original states, not the errors, dominates -- then the claimed convergence of the residual to zero in the absence of attack is unsupported. Equivalently, run the attack-free simulation with the nominal-model mismatch of (47) and measure whether the residual $r$ converges to zero; a residual that stays bounded away from zero would falsify the detection claim.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that a robust zero dynamics attack -- a stealthy input signal that drives the unobservable zero dynamics unstable while leaving the measured output nearly unchanged -- can be detected and neutralized in a general MIMO nonlinear system using only a nominal model. The paper defines the residual $r := u_c + \alpha - u_{mc}$, the difference between the actual plant input (controller output plus attack) and the controller output of a parallel nominal model; Theorem 1 claims this residual converges to zero when no attack is present and to the attack signal when one is, provided the closed-loop Lyapunov decay rate dominates the Lipschitz and gain constants ($c_3 > 6c_5c_4$). Theorem 2 then claims that, under Assumptions 1-3 and the parameter conditions $c_3 \geq c_4c_5(3+c_5c_8)+\lambda$ and $\lambda \leq c_4/(c_7c_9)$, the neural-network estimate $z_r = G_n^{-1}(F_n(w,x^*))$ keeps the estimation error $\alpha - z_r$ bounded, the zero dynamics bounded, the output tracking the reference, and the closed loop stable. The four-tank simulations are presented as evidence that the detector reacts at the attack instant and that the recovery loop keeps the zero dynamics from collapsing.

Load-bearing premise

The argument depends on the assumption that a single stability certificate for the original closed-loop system also works for the difference between the real system and the nominal model, even though that difference includes extra mismatch terms the original system does not have.

Editorial extensions

If this is right

  • The detection stage can be added to an existing closed loop without modifying the controller, because the residual monitor runs in parallel and uses only the plant input and the nominal model's controller output.
  • Recovery needs no pre-collected attack data: the neural-network weights are updated online from the residual, so the method is intended to work against previously unseen attack signals.
  • Because the nominal model is allowed to differ from the plant, the scheme is claimed to work under parametric and structural uncertainty, as well as measurement noise.
  • The comparison scenario indicates that the proposed recovery keeps the zero dynamics near their nominal values while the baseline method's attack success rate stays near 100 percent.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The residual-amplitude detector would also respond to large model mismatch or disturbances in the absence of an attack; a practical implementation would need a threshold or a test that separates 'residual due to mismatch' from 'residual due to attack', a design choice the paper leaves open.
  • A natural extension is to quantify the maximum zero-dynamics excursion during the time between attack injection and detection, since recovery is only switched on after detection and Theorem 2 does not bound the transient before then.
  • The proof structure suggests a testable sufficient condition: if the nominal model is close enough to the plant that the combined Lipschitz and gain constants satisfy $c_3 > 6c_5c_4$, the detector should be immune to mismatch-induced false alarms; constructing a plant-model pair on the boundary of this inequality would test the margin of the result.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes a two-stage strategy for detecting and mitigating robust zero dynamics attacks in MIMO nonlinear control systems. In the detection stage, a residual signal is formed as the difference between the actual plant input and the output of a nominal-model controller driven by the plant output; the residual is claimed to converge to zero without attack and to the attack signal under attack. In the mitigation stage, a neural network with online weight adaptation estimates the attack signal from the system states and the residual, and the estimate is subtracted from the plant input to recover nominal closed-loop behavior. The claims are supported by Lyapunov-based theorems and by simulations on a four-tank system with constant, noisy, and time-varying references.

Significance. If the main theorems were correct, the paper would offer a notable advance: detection and mitigation of robust zero dynamics attacks for general MIMO nonlinear systems without an exact model and without offline training data. The simulation suite is appropriate and includes noise and time-varying references, and the comparison with a prior resilient controller is informative. However, the central analytical claims are not rigorously established. The proof of Theorem 1 applies a Lyapunov function of the true closed-loop system to an error system with different vector fields and derives an inequality that does not imply convergence. The proof of Theorem 2 contains a dimensionally inconsistent equality and introduces an ad hoc auxiliary variable whose behavior is not analyzed. These issues are load-bearing: the detection threshold, the adaptation error, and the recovery cancellation all depend on the unproven convergence and boundedness statements. The paper's strongest advertised contribution is therefore not supported by the presented analysis.

major comments (4)
  1. [Section III, Eqs. (14)-(17)] The proof of Theorem 1 states that because Eq. (14) is similar to the system structure in (1), Assumption 3 allows V_sigma to serve as a Lyapunov function for the error dynamics. This is not justified: Assumption 3 gives inequalities for V_sigma along the true closed-loop vector field F, whereas the error vector field in (14) contains the perturbation terms Delta1, Delta2, Delta3 and the extra coupling G_sigma(u_c - u_mc). Even if (17) were obtained, it reads ||V_dot_sigma|| <= -c3 ||[z_tilde,x_tilde,s_tilde]||^2 + 6 c5 c4 ||[z,x,s]||^2, which mixes the error norm with the true-state norm. Since ||[z,x,s]|| is bounded but not necessarily decaying, this inequality implies at most boundedness of the error; it does not imply convergence of z_tilde, x_tilde, s_tilde to zero, nor does it imply that the residual r in (9) converges to zero without attack. The detection claim and the subsequent use of r as the adaptation error e in (18) rest on this unproven convergence, so the detection guarantee is unsupported.
  2. [Section IV, Eq. (37)] The equality -||(alpha - z_r)^T (I_w G)|| = -lambda ||(alpha - z_r)^T (alpha - z_r) + (alpha - z_r)^T (u_c - u_mc)|| is not valid. From (18), G = lambda (dG_n^{-1}F_n/dw) e / ||dG_n^{-1}F_n/dw||^2 and I_w = dG_n^{-1}F_n/dw, so the product I_w G is not equal to lambda e unless the Jacobian term reduces to a projection and e coincides with (alpha - z_r) + (u_c - u_mc). The proof of boundedness of alpha - z_r depends on this equality and the subsequent completion of the square in (38), so the Lyapunov argument for Theorem 2 does not go through.
  3. [Section IV, Eqs. (28)-(29) and (43)] The auxiliary variable L is introduced in (28) with dynamics (29) purely to absorb cross terms in the Lyapunov derivative, but the proof never establishes that L remains bounded on the time interval of interest. Moreover, Eq. (43) asserts the inequality L^T L_dot + ||(alpha - z_r)^T (-I_w H_n(...) + I_delta H_n(...))|| <= 0 without derivation from (29). The stability and boundedness conclusions of Theorem 2 rely on this cancellation, so the proof is incomplete.
  4. [Section II, Eq. (2) and Section IV, Eqs. (18)-(19)] The attack model in (2) and the neural-network estimator in (18)-(19) use the same nominal functions F_n, G_n^{-1}, and H_n: the attack signal is defined through G_n^{-1}(F_n(delta+z_n^*,x*) - F_n(z^*,x*)), and the recovery signal is G_n^{-1}(F_n(w,x*)) with an adaptation law built from the same G_n^{-1} and F_n. Consequently, the estimator is structurally matched to the attack generator, which is a form of design-by-model rather than independent estimation of a general attack. The claim in the abstract that the method works 'without relying on an accurate dynamical model' is not supported for attack signals not generated in this form, and the paper provides no analysis for general attack signals.
minor comments (4)
  1. [Section III, Eq. (14)] The definition of Delta3 in (10) uses the argument M(x - x_mp, y_ref - x, s - s_mc), while the controller in (8) is written as M(x_mp, x, s_mc); the notation is inconsistent and should be clarified.
  2. [Theorem 1 statement] The theorem states that the residual 'converges to the attack signal in the presence of the attack;' the proof only argues that the residual becomes nonzero when the attack occurs, not that it converges to alpha. This wording should be corrected to match the argument.
  3. [Section IV, Eqs. (41)-(42)] The transition from (41) to (42) is unexplained: the term G_n^{-1}(F_n(delta+z_n^*,x*) - F_n(z^*,x*)) is replaced by r in (42), but r was defined in (9) as u_c + alpha - u_mc, not as the inverse-image expression. The substitution should be justified.
  4. [Throughout] There are several typos and grammatical issues, including 'Lyapanuv' in Remark 4 and 'the following two conditions simultaneously' in Section IV; a careful proofreading pass is needed.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the residual detector and neural-network estimator share the nominal model with the attack model by design, but no claimed result reduces by construction to its inputs, and the cited attack/Lyapunov results are external prior work, not self-citations.

full rationale

The paper's derivation chain does not contain a load-bearing step that is equivalent to its own inputs by definition or by self-citation. The attack model in Eq. (2) and the neural-network estimator in Eqs. (18)-(19) both use the nominal functions Fn, Gn, and Hn, but this is a design choice: the estimator's output zr = Gn^{-1}(Fn(w,x*)) is not equal to the attack signal alpha by construction, and the adaptation law updates w from the residual r = uc + alpha - umc, which is a standard prediction-error formulation rather than a fitted parameter being renamed as a prediction. Theorem 1 attempts to prove convergence of r to zero (no attack) and to alpha (attack) from the Lyapunov/Assumption structure, and Theorem 2 attempts to prove boundedness of the estimation error from Assumptions 1-3 plus parameter conditions. These are substantive derivations, even if they contain proof gaps. In particular, the step in Theorem 1 where V_sigma is reused as a Lyapunov function for the error dynamics (14) and the step in Theorem 2 where the auxiliary variable L is chosen to cancel terms are correctness concerns, not circular reductions: the conclusions are not obtained by restating the assumptions. The cited results [6] and [25] on robust zero dynamics attacks and the four-tank system's Lyapunov property are external works by different authors, so there is no self-citation chain carrying the argument. The simulation results are also compared against an external baseline. No circular step of the enumerated kinds can be exhibited with a specific equation-to-equation reduction.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central results rest on Assumptions 1-3 and on unstated choices for lambda and initial weights. The 'neural network' is not a learned function but a parametrized copy of the nominal model. No new physical or conceptual entities are introduced; the auxiliary variable L is a proof device, not an entity.

free parameters (3)
  • Learning rate lambda = not specified
    Used in the weight update law (18). The paper states conditions involving lambda but does not give a concrete value or tuning procedure.
  • Initial neural network weights w(0) = not specified
    The adaptation law (18) requires initial weights. No default or heuristic is provided.
  • Detection threshold for residual amplitude = not specified
    The detection method 'evaluates the amplitude' of the residual, but no threshold or statistical test is defined.
assumptions (4)
  • domain assumption Assumption 1: System and model functions H_sigma, F_sigma, G_sigma, M, Hn, Fn, Gn are Lipschitz with common bound c5.
    Used repeatedly in Theorem 1 and Theorem 2 to bound the perturbation terms Delta1-Delta5. Section III.
  • domain assumption Assumption 2: Gn^-1 and Fn satisfy bi-Lipschitz conditions with constants c7, c8, c9.
    Required to bound ||w - delta_tilde|| in the recovery proof. Section III.
  • domain assumption Assumption 3: There exists a Lyapunov function V_sigma satisfying (4) for the attack-free closed-loop system, and c4 is known or bounded.
    This is the backbone of both theorems. The paper argues it is simple to find for an existing controller, but no constructive method is given. Section III.
  • ad hoc to paper The nominal model (Fn, Gn, Hn) is known and its functions can be evaluated with the current weights w as inputs.
    The neural network is defined as compositions of Fn and Gn^-1, which presupposes exact knowledge of these functions and their derivatives for the update law. This is stronger than 'no accurate model' in practice.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Online Detection and Mitigation of Robust Zero Dynamics Anomaly Behavior in MIMO Nonlinear Control Systems." pith.science (2026). https://pith.science/paper/SRI66C2K

@misc{pith2026250603298,
  author       = {Pith},
  title        = {Pith review of: Online Detection and Mitigation of Robust Zero Dynamics Anomaly Behavior in MIMO Nonlinear Control Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SRI66C2K}},
  note         = {Machine review of arXiv:2506.03298}
}
read the original abstract

This paper presents a methodology to detect robust zero dynamics anomaly behavior and mitigate the impacts in general multi-input multi-output (MIMO) nonlinear systems. The proposed method guarantees the resiliency and stability of the closed-loop system without relying on an accurate dynamical model. The presented method operates in two stages. First, it measures the difference between the system input and that of the model as a residual signal to detect the anomaly behavior. After detecting the attack, a recovery signal is generated to restore the system to its nominal condition. In this stage, a neural network model is used to estimate the anomaly signal and recover the closed-loop system. The weights of the neural network model are updated online using adaptation rules without needing prior data for training. The accuracy and performance of the proposed methods are verified by simulating various scenarios on a fourtank system.

Figures

Figures reproduced from arXiv: 2506.03298 by the authors.

Figure 1
Figure 1. The general structure of the proposed robust zero dynamics attack [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. The general structure of the proposed recovery algorithm. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 1
Figure 1. The objective of the system controller is enabling the system to track a reference signal with a bounded error while all state variables remain above zero. The dynamics of the four-tank system defined in (44) and (45) include constant values and radical functions. It can easily be shown that these terms satisfy both Lipschitz and bi-Lipschitz conditions (Assumptions 1 and 2) over any closed interval greater than zer… view at source ↗
Figures from the paper (10 more)
Figure 4
Figure 4. Figure 4: Model output in scenario A: The output of the model with a constant reference. The residual signals are shown in Fig. (5). When the attack occurs, the residual signals become nonzero. In contrast, residual signals converge to zero during steady states and in the absenc…
Figure 5
Figure 5. Figure 5: Residuals in scenario A: Residual signals used to detect the occurrence of the attack and update the neural network weights. 0 100 200 300 400 500 600 700 800 900 1000 Time (s) -2 -1 0 1 2 3 Voltage (V) First difference Second difference [PITH_FULL_IMAGE:figures/full_…
Figure 6
Figure 6. Figure 6: Difference between residuals and attack signals in scenario A: It can be seen that after a short transients, residuals converge to attack signals. success rate for this scenario are zero, validating the recovery and detection algorithm’s acceptable performance. 0 100 2…
Figure 7
Figure 7. Figure 7: Zero dynamics the system in scenario A: Zero dynamics of the system stay stable under the proposed recovery algorithm. 0 100 200 300 400 500 600 700 800 900 1000 Time (s) 0 50 100 Height (cm) First zero dynamics Second zero dynamics [PITH_FULL_IMAGE:figures/full_fig_p…
Figure 10
Figure 10. Figure 10: System output in scenario B: The output of the closed-loop system under noisy measurements. 0 200 400 600 800 1000 1200 Time (s) 60 70 80 90 100 Height (cm) First zero dynamics Second zero dynamics [PITH_FULL_IMAGE:figures/full_fig_p009_10.png]
Figure 11
Figure 11. Figure 11 [PITH_FULL_IMAGE:figures/full_fig_p009_11.png]
Figure 12
Figure 12. Figure 12: Residual signals in scenario B: Residual signals used to detect the occurrence of the attack and update the neural network weights. 600 650 700 750 800 850 900 950 1000 Time (s) -10 -5 0 5 Voltage (V) First output Second output [PITH_FULL_IMAGE:figures/full_fig_p010_…
Figure 16
Figure 16. Figure 16: The neural network output in scenario C: Recovery signals are used to restore the closed-loop system to its normal operating point irrespective of the properties of the reference signal. method cannot stabilize the zero dynamics of the system in the presence of a robu…
Figure 17
Figure 17. Figure 17: Zero dynamics the system in scenario D: Zero dynamics of the system cannot show a desirable behavior under the proposed recovery algorithm in [26]. VI. CONCLUSION In this work, a new detection and stabilization strategy for general MIMO nonlinear cyber-physical system…
Figure 14
Figure 14. Figure 14: System output in scenario C: The output of the closed-loop system with a time-varying reference signal. D. Comparison Study In [26], a resilient control strategy is presented to recover nonlinear affine systems under actuator attack. Fig. (17) shows the zero dynamics …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

26 extracted references · 26 canonical work pages

  1. [1]

    An ensemble deep learning-based cyber- attack detection in industrial control system,

    A. Al-Abassi, H. Karimipour, A. Dehghantanha, and R. M. Parizi, “An ensemble deep learning-based cyber- attack detection in industrial control system,” IEEE Ac- cess, vol. 8, pp. 83 965–83 973, 2020

    work page 2020

  2. [2]

    Covert attacks in cyber-physical control sys- tems,

    A. O. de S ´a, L. F. R. da Costa Carmo, and R. C. Machado, “Covert attacks in cyber-physical control sys- tems,” IEEE Transactions on Industrial Informatics , vol. 13, no. 4, pp. 1641–1651, 2017

    work page 2017

  3. [3]

    Denial of service attacks on network-based control systems: impact and mitigation,

    M. Long, C.-H. Wu, and J. Y . Hung, “Denial of service attacks on network-based control systems: impact and mitigation,” IEEE Transactions on Industrial Informatics, vol. 1, no. 2, pp. 85–96, 2005

    work page 2005

  4. [4]

    Zero- dynamics attack, variations, and countermeasures,

    H. Shim, J. Back, Y . Eun, G. Park, and J. Kim, “Zero- dynamics attack, variations, and countermeasures,” in Security and Resilience of Control Systems: Theory and Applications. Springer, 2022, pp. 31–61

    work page 2022

  5. [5]

    Stealthy adversaries against uncertain cyber-physical systems: Threat of robust zero-dynamics attack,

    G. Park, C. Lee, H. Shim, Y . Eun, and K. H. Johansson, “Stealthy adversaries against uncertain cyber-physical systems: Threat of robust zero-dynamics attack,” IEEE Transactions on Automatic Control , vol. 64, no. 12, pp. 4907–4919, 2019

    work page 2019

  6. [6]

    Stealthiness of zero- dynamics attacks against uncertain nonlinear systems : A case study with quadruple-tank process,

    G. Park, C. Lee, and H. Shim, “Stealthiness of zero- dynamics attacks against uncertain nonlinear systems : A case study with quadruple-tank process,” 23rd Interna- tional Symposium on Mathematical Theory of Networks and Systems Hong Kong University of Science and Tech- nolog, 2018

    work page 2018

  7. [7]

    Revealing stealthy attacks in control systems,

    A. Teixeira, I. Shames, H. Sandberg, and K. H. Jo- hansson, “Revealing stealthy attacks in control systems,” 2012 50th Annual Allerton Conference on Communica- tion, Control, and Computing (Allerton) , pp. 1806–1813, 2012

    work page 2012

  8. [8]

    A data-driven attack detection approach for dc servo motor systems based on mixed optimization strategy,

    X. Li and X.-Y . Shen, “A data-driven attack detection approach for dc servo motor systems based on mixed optimization strategy,” IEEE Transactions on Industrial Informatics, vol. 16, pp. 5806–5813, 2020

    work page 2020

Show all 26 references
  1. [9]

    Reliable control policy of cyber- physical systems against a class of frequency-constrained sensor and actuator attacks,

    X. Huang and J. Dong, “Reliable control policy of cyber- physical systems against a class of frequency-constrained sensor and actuator attacks,” IEEE Transactions on Cy- bernetics, vol. 48, no. 12, pp. 3432–3439, 2018

    2018

  2. [10]

    An active detection strategy based on dimensionality reduction for false data injection attacks in cyber-physical systems,

    M. Attar and W. Lucia, “An active detection strategy based on dimensionality reduction for false data injection attacks in cyber-physical systems,” IEEE Transactions on Control of Network Systems , vol. 10, no. 4, pp. 1844– 1854, 2023

    2023

  3. [11]

    Passive attack detection for a class of stealthy intermittent integrity attacks,

    K. Zhang, C. Keliris, T. Parisini, B. Jiang, and M. M. Polycarpou, “Passive attack detection for a class of stealthy intermittent integrity attacks,” IEEE/CAA Jour- nal of Automatica Sinica , vol. 10, no. 4, pp. 898–915, 2023

    2023

  4. [12]

    Modeling and detection scheme for zero-dynamics at- tack on wind power system,

    Z. Wang, H. Zhang, X. Cao, E. Liu, H. Li, and J. Zhang, “Modeling and detection scheme for zero-dynamics at- tack on wind power system,” IEEE Transactions on Smart Grid , vol. 15, no. 1, pp. 934–943, 2024

    2024

  5. [13]

    Attack detection design for dc microgrid using eigenvalue assignment approach,

    S. Tan, P. Xie, J. M. Guerrero, J. C. Vasquez, Y . Li, and X. Guo, “Attack detection design for dc microgrid using eigenvalue assignment approach,” Energy Reports, vol. 7, pp. 469–476, 2021

    2021

  6. [14]

    Monitor- ing and detection of malicious adversarial zero dynamics attacks in cyber-physical systems,

    A. Baniamerian, K. Khorasani, and N. Meskin, “Monitor- ing and detection of malicious adversarial zero dynamics attacks in cyber-physical systems,” 2020 IEEE Confer- ence on Control Technology and Applications (CCTA) , pp. 726–731, 2020

    2020

  7. [15]

    Neutralizing zero dynamics attack on sampled- data systems via generalized holds,

    J. Kim, J. Back, G. Park, C. Lee, H. Shim, and P. G. V oul- garis, “Neutralizing zero dynamics attack on sampled- data systems via generalized holds,” Automatica, vol. 113, p. 108778, 2020

    2020

  8. [16]

    Zero-dynamics attack on wind turbines and countermeasures using generalized hold and generalized sampler,

    D. Kim, K. Ryu, and J. Back, “Zero-dynamics attack on wind turbines and countermeasures using generalized hold and generalized sampler,” Applied Sciences, vol. 11, no. 3, 2021

    2021

  9. [17]

    A survey of deep neural network architectures and their applications,

    W. Liu, Z. Wang, X. Liu, N. Zeng, Y . Liu, and F. E. Alsaadi, “A survey of deep neural network architectures and their applications,” Neurocomputing, vol. 234, pp. 11–26, 2017

    2017

  10. [18]

    Unsupervised random forest: a tutorial with case studies,

    N. L. Afanador, A. Smolinska, T. N. Tran, and L. Blanchet, “Unsupervised random forest: a tutorial with case studies,” Journal of Chemometrics , vol. 30, no. 5, pp. 232–241, 2016

    2016

  11. [19]

    De- tection of sensor attack and resilient state estimation for uniformly observable nonlinear systems having redun- dant sensors,

    J. Kim, C. Lee, H. Shim, Y . Eun, and J. H. Seo, “De- tection of sensor attack and resilient state estimation for uniformly observable nonlinear systems having redun- dant sensors,” IEEE Transactions on Automatic Control , vol. 64, no. 3, pp. 1162–1169, 2019

    2019

  12. [20]

    Relative information loss in the pca,

    B. C. Geiger and G. Kubin, “Relative information loss in the pca,” 2012 IEEE Information Theory Workshop , pp. 562–566, 2012

    2012

  13. [21]

    Detection of sensor attack and resilient state estimation for uniformly observable nonlinear systems,

    J. Kim, C. Lee, H. Shim, Y . Eun, and J. H. Seo, “Detection of sensor attack and resilient state estimation for uniformly observable nonlinear systems,” 2016 IEEE 55th Conference on Decision and Control (CDC) , pp. 1297–1302, 2016

    2016

  14. [22]

    Pinning controllability of complex network systems with noise,

    D. A. Burbano-L, G. Russo, and M. Di Bernardo, “Pinning controllability of complex network systems with noise,” IEEE Transactions on Control of Network Systems, vol. 6, no. 2, pp. 874–883, 2018

    2018

  15. [23]

    Robust nonlinear observer-based controller design for one-sided lipschitz switched systems with time-varying delays,

    W. Iqbal, I. Ghous, E. A. Ansari, Z. Duan, M. Imran, A. A. Khan, and M. T. Humayun, “Robust nonlinear observer-based controller design for one-sided lipschitz switched systems with time-varying delays,” Journal of the Franklin Institute , vol. 360, no. 3, pp. 2046–2067, 2023

    2023

  16. [24]

    Reinforcement learning based- adaptive tracking control for a class of semi-markov non- lipschitz uncertain system with unmatched disturbances,

    Z. Wang and J. Liu, “Reinforcement learning based- adaptive tracking control for a class of semi-markov non- lipschitz uncertain system with unmatched disturbances,” Information Sciences , vol. 626, pp. 407–427, 2023

    2023

  17. [25]

    When adversary encounters uncertain cyber-physical systems: Robust zero-dynamics attack with disclosure resources,

    G. Park, H. Shim, C. Lee, Y . Eun, and K. H. Johansson, “When adversary encounters uncertain cyber-physical systems: Robust zero-dynamics attack with disclosure resources,” in 2016 IEEE 55th Conference on Decision and Control (CDC) , 2016, pp. 5085–5090

    2016

  18. [26]

    Cloud-based underactuated resilient control for cyber- physical systems under actuator attacks,

    Y . Zhao, C. Zhou, Y .-C. Tian, J. Yang, and X. Hu, “Cloud-based underactuated resilient control for cyber- physical systems under actuator attacks,” IEEE Transac- tions on Industrial Informatics , vol. 19, no. 5, pp. 6317– 6325, 2023

    2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.