betaFeedback

WorksEngineering

Balancing Confidentiality and Attack Detectability in Networked Control Systems over Lossy Wireless Channels

Abstract

We study the joint protection of a networked control system (NCS) against eavesdropping and data integrity attacks over a lossy wireless channel. The eavesdropper intercepts the transmitted sensor measurements and control inputs to reconstruct the system state. In contrast, the integrity attacker in this study replaces the sensor measurements with fabricated data to drive the system to undesired operating conditions. Although these two threats are typically studied separately, a wireless NCS should be protected against both simultaneously, and a defence against one affects the other. We employ an observation-encoding scheme that renders the eavesdropper's state-estimation error covariance unbounded under packet drops, and show that transmission energy governs how long this error remains bounded. In particular, lower transmission energy improves confidentiality. For integrity protection, we combine stochastic physical watermarking with a CUSUM-based sequential change detector. The proposed attack detection scheme accounts for intermittent communication and encoded measurements. The detection delay is governed by the Kullback–Leibler divergence (KLD) between the pre- and post-attack distributions of the test signal, and we show that a higher transmission energy increases the KLD and thereby reduces the detection delay. Furthermore, we derive analytical expressions for the eavesdropper's estimation error and the worst-case expected KLD, providing guidelines for selecting transmission energy to balance confidentiality and integrity protection and characterising the privacy–security trade-off. Finally, we run numerical simulations to validate our theoretical findings and demonstrate the privacy–security trade-off in practical scenarios.

*This work was fully supported by the Wallenberg AI, Autonomous Systems and Software Program (WASP) funded by the Knut and Alice Wallenberg Foundation.

INTRODUCTION

Networked control systems (NCSs), in which sensing, computation and actuation are connected through communication networks, are common in several applications such as power systems, transportation and industrial automation [1, 2, 3]. Wireless networks make such systems flexible and inexpensive to deploy, but they also introduce challenges such as channel unreliability, packet drops, interference, and energy limitations, which degrade estimation and control performance [4], [5]. Because most NCSs use off-the-shelf networking components, commodity software, and low-cost sensors with limited energy and computational capabilities, they are also prone to attacks. Attacks on NCSs can be broadly classified as passive or active. Passive attacks seek to extract information without modifying transmitted data or disrupting system operation. For example, an eavesdropper may intercept sensor measurements and control inputs and use knowledge of the system dynamics to estimate sensitive states [6], [7]. Unencrypted communication can facilitate such inference. Although limited energy, computational resources, and stringent latency requirements can complicate the deployment of encryption [8]. On the other hand, active attacks manipulate data or disrupt system operation. These include data deception attacks, in which an adversary falsifies measurements or control commands to mislead the controller or alter the physical process. A replay attack is a form of data deception in which previously recorded, legitimate data are retransmitted as current data, potentially concealing malicious activity. For example, Stuxnet attack targeted uranium-enrichment centrifuges at Iran’s Natanz facility by manipulating valves that controlled gas flow while replaying recorded values from normal operation to conceal the manipulation. Such attacks can degrade performance, and drive the system outside safe operating limits.

Existing studies often address eavesdropping and data deception attacks separately. However, a defense against one threat can influence vulnerability to the other, creating potential trade-offs between privacy protection and attack resilience, as studied in [9], [10]. These interactions motivate joint defenses that address both threats while maintaining acceptable control performance, particularly in safety-critical applications. Developing such defenses requires characterizing their achievable performance bounds and understanding how design parameters influence the protection provided against each threat. Motivated by these challenges, we studied a joint defense mechanism against eavesdropping and data deception attacks in NCSs. Specifically, we consider measurement replacement attacks, a form of data deception in which an adversary spoofs sensor measurements by replacing legitimate data with falsified values.

State Privacy. Traditional cryptography-based methods can protect data confidentiality but can impose computational and energy costs on resource-constrained devices [11], [12]. This motivates approaches based on system dynamics and communication properties, including state encoding, transmission scheduling, and noise injection.

State-secrecy code-based approaches encode the current state or its local estimate using the information previously received by the legitimate estimator. When the eavesdropper misses a packet that is received by the legitimate estimator, it loses reference information needed to decode subsequent transmissions [13]. Related schemes address acknowledgement manipulation [14], operation without acknowledgements [15], active eavesdroppers [16], and multi-sensor networks [17]. In a different approach, transmission scheduling controls when information is sent to balance legitimate estimation accuracy and confidentiality [18, 19]. Furthermore, noise injection can obscure transmitted information [20], although it may also degrade legitimate estimation. With suitable channel knowledge and spatial dimensions, artificial noise can be placed in the null space of the legitimate channel [21], avoiding additional interference at that legitimate receiver while impairing eavesdropping [11]. In our study, we build on this state-encoding idea and apply it to encode the remote sensors’ observations sent to the state estimator.

Data Deception Attacks. Data deception attacks on the cyber-physical systems, for example NCS, where an adversary manipulates sensor measurements and/or control signals, have been studied extensively in the literature [22, 23]. The ability of such an adversary to compromise a system within a limited number of sensor manipulations, where measurements can be replaced, deleted, or inserted, has also been characterized formally in [24]. Since detectors that only observe the received measurements may fail to detect such attacks [23, 25], active schemes based on physical watermarking add a private random signal to the control input. Therefore, test signals, for instance innovation signal from a Kalman filter, not generated by the true plant then lack the expected correlation with the watermark, which changes the innovation statistics [25, 26, 27, 28]. Alternatively, parity-space residuals have been shown to capture the effect of replayed input and output data and to distinguish it from system faults, with both passive and active designs to enhance detection [29]. Recent work has reduced the cost of watermarking, e.g., through event-based [30] or parsimonious [23] policies, and has combined it with moving-target defense [31] or dynamic encryption [32]. Moving-target defense has further been extended to a data-driven setting, where a time-domain estimate of the ν-gap metric is used to separate stealthy attacks from faults in [33]. Additionally, countermeasures that exploit asynchronous sampling have been proposed against stealthy attacks on aperiodic sampled-data systems in [34].

To detect attacks with minimal delay, sequential change-detection tests such as the cumulative sum (CUSUM) procedure [35] have been combined with watermarking [23, 36, 37]. In particular, [23] derived the conditional innovation distributions before and after a measurement replacement attack and used them in an optimal CUSUM test. These results, however, assume that all measurements are received, so that the Kalman filter operates in steady state and the innovation is stationary. Attack detectability and identifiability with asynchronous, delayed, and multi-rate measurements have been analyzed using a lifted system model [38], and watermarking under packet drops has been considered in [39]; however, neither has been combined with encoding or sequential detection.

Combining these methods with measurement encoding and sensor packet losses requires further analysis. Packet losses make the estimation error covariance depend on the reception history [4]. Encoding also links each decoded measurement to the last received packet, which may already be falsified. Therefore, an attack detector, which is combined with a defence mechanism against eavesdropping must account for both the reception history and the encoding dependencies.

Our Contributions

In this paper, we study the joint protection of an NCS against eavesdropping and measurement replacement attacks over a lossy sensor–estimator link. To this end, we combine observation encoding and stochastic physical watermarking to enhance both confidentiality and attack detection. In the proposed scheme, the sensors encode their observations before transmission based on the last observations received by the legitimate user, and the legitimate user decodes them accordingly. Since the channel is lossy, the eavesdropper may miss a packet that the legitimate user receives, causing the eavesdropper to lose the reference needed for decoding. We refer to such events as critical events. Once such an event occurs, the eavesdropper’s state-estimation error covariance grows without bound, while the legitimate estimator’s error covariance remains bounded. Therefore, it is important to reduce the average time before a critical event occurs to enhance the overall confidentiality of the system. Our study shows how the transmission energy affects the time to the critical event, provided that the other factors influencing the channel quality remain constant. Note that, similar to several previous studies, our proposed defense mechanism does not require the system to be open-loop unstable to be effective. On the other hand, to detect measurement replacement attacks effectively, we design a CUSUM-based sequential detection scheme that uses the known watermark. To this end, we derive the conditional innovation distributions under nominal and attacked conditions, taking into account packet losses and encoding, and use them to update the CUSUM statistic only when a packet is received. To study the performance of the detector, we derive an analytical expression for the worst-case expected Kullback–Leibler divergence (KLD) between the nominal and attacked distributions, which serves as a measure of attack detectability. In other words, a lower KLD means a longer detection delay, which is undesirable. Furthermore, our study shows that the higher the transmission energy, the longer the average time to the critical event but the higher the attack detectability, highlighting the trade-off between confidentiality and detection performance. Our main contributions are the following:

  1. We combine observation encoding and stochastic physical watermarking to jointly address eavesdropping and measurement replacement attacks in networked control systems over lossy wireless channels.

  1. We show that, after a critical event, the encoding makes the eavesdropper’s state-estimation error covariance unbounded while keeping the legitimate estimator’s

error covariance bounded. We show how transmission energy affects the time to this critical event.

  1. We derive the distributions of the innovation signal under nominal and attacked conditions under packet losses and encoding, and design a CUSUM-based sequential attack detection scheme. We also derive an analytical expression for the worst-case expected KLD. These results guide transmission-energy selection to balance confidentiality and attack detection.

  1. Numerical simulations support the analysis and illustrate the confidentiality–detection trade-off.

Organization of the Paper

The paper is organized as follows. Section II presents the system and threat models. Section III introduces the encoding scheme and the watermarking-based CUSUM detector. Section IV analyzes their performance and the resulting confidentiality–detection trade-off. Numerical results are reported in Section V, and Section VI concludes the paper.

Notation

We use the following notation throughout the paper. Vectors are column vectors and are denoted by bold lowercase letters, e.g., x\mathbf{x}. Matrices are denoted by bold uppercase letters, e.g., A\mathbf{A}. The transpose of a matrix A\mathbf{A} is denoted by A⊤\mathbf{A}^{\top}. The identity matrix of appropriate dimension is denoted by I\mathbf{I}. The expectation operator is denoted by E[⋅]\mathbb{E}[\cdot]. The trace of a matrix A\mathbf{A} is denoted by tr⁡(A)\operatorname{tr}(\mathbf{A}). The spectral radius of a matrix A\mathbf{A} is denoted by ρ(A)\rho(\mathbf{A}). The notation A⪰0\mathbf{A} \succeq0 (A≻0\mathbf{A} \succ0) indicates that A\mathbf{A} is positive semidefinite (positive definite). The notation N(m,P)\mathcal{N}(\mathbf{m}, \mathbf{P}) denotes a Gaussian distribution with mean m\mathbf{m} and covariance P\mathbf{P}. ∣⋅∣|\cdot| denotes the determinant of a matrix.

Preliminaries: System and Threat models

In this section, we formulate the problem of secure state estimation and attack detection in NCS with lossy communication channels. We describe the system dynamics, the communication model, the encoding scheme, and the attack model.

In our study we have considered a time-invariant linear stochastic system for modeling the plant dynamics, see Subsection II-A. The sensor measurements are transmitted over a lossy communication channel to a remote estimator, which reconstructs the system state. The receiver (legitimate estimator) sends acknowledgements to the sensor. We further assume, the acknowledgements are reliably received by the sensor, since the controller–estimator block generally has much higher energy resources for communication. The estimator reconstructs the system state based on the received packets. The estimated states are then used for feedback control. The controller generates the control input based on the estimated state and transmits it to the actuator over a wireless network. An illustration of this setup is shown in Figure 1 along with the eavesdropper and the attacker. We assume the controller has much higher energy resources compared to the sensor, allowing it to reliably transmit control signals. In other words, controller–actuator communication is assumed to be effectively lossless. Therefore, the primary focus of this work is on securing the sensor–controller communication against eavesdropping and attacks. A more general study considering attacks on the controller–actuator link, which is also lossy, is left for future work.

Figure 1. Scheme 1 with state encoding. The encoder sends zt=yt−Mt−tkytkz_t = y_t - M_{t-t_k}y_{t_k} over a lossy channel (δt\delta_t), where an attacker may replace it with ztaz_t^a. The decoder reconstructs ytdy_t^d for the Kalman filter, and the acknowledgement keeps both memories in sync.

Plant Model

We consider a linear stochastic system for our study described by the following state-space equations

xt+1=Axt+But+wt,(1)\mathbf{x}_{t+1} = \mathbf{A}\mathbf{x}_t + \mathbf{B}\mathbf{u}_t + \mathbf{w}_t, \tag*{(1)}
yt=Cxt+vt,(2)\mathbf{y}_t = \mathbf{C}\mathbf{x}_t + \mathbf{v}_t, \tag*{(2)}

where xt∈Rn\mathbf{x}_t \in\mathbb{R}^n is the state, ut∈Rm\mathbf{u}_t \in\mathbb{R}^m is the control input and yt∈Rp\mathbf{y}_t \in\mathbb{R}^p is the measurement.

Assumption 1 (Noises): wt∼N(0,Q)\mathbf{w}_t \sim\mathcal{N}(0,\mathbf{Q}), vt∼N(0,R)\mathbf{v}_t \sim\mathcal{N}(0,\mathbf{R}) with Q⪰0\mathbf{Q} \succeq0, and R≻0\mathbf{R} \succ0 are i.i.d. in time, mutually independent and independent of x0∼N(x^0∣0,P0)\mathbf{x}_0 \sim\mathcal{N}(\hat{\mathbf{x}}_{0|0},\mathbf{P}_0).

Lossy channel model

As noted earlier, we assume only the sensor–estimator link is subject to packet losses. Let δt∈{0,1}\delta_t \in\{0,1\} indicate successful packet reception at time tt, with δt=1\delta_t = 1 if the packet is received and δt=0\delta_t = 0 otherwise. We model δt\delta_t as a Bernoulli random variable with arrival probability δˉ\bar{\delta}. Following [40], we consider quadrature amplitude modulation (QAM), and for a fixed transmit energy E≥0E \ge0. Under this transmission scheme, δˉ\bar{\delta} is related to EE as

δˉ=1−ϵE,ϵ=exp⁡(−αN0Wbw)∈(0,1),(3)\bar{\delta} = 1 - \epsilon^E,\qquad\epsilon= \exp\left(-\frac{\alpha}{N_0 W_{\mathrm{bw}}}\right) \in(0,1), \tag*{(3)}

where N0N_0 is the noise power spectral density, WbwW_{\mathrm{bw}} is the channel bandwidth, and α∈(0,1)\alpha\in(0,1) is a modulation-dependent constant. Thus, achieving a target arrival probability δˉ∈[0,1)\bar{\delta} \in[0,1) requires the transmit energy

E(δˉ)=ln⁡(1−δˉ)ln⁡ϵ.E(\bar{\delta}) = \frac{\ln(1-\bar{\delta})}{\ln\epsilon}.

We assume that the transmit energy is held fixed across transmissions and does not depend on the data. We further assume that the packet-reception process is independent of the transmitted content.

Estimator and Controller

The legitimate receiver employs a state estimator which predicts the state using the known system model and updates this prediction whenever a measurement is received. Under nominal operation, the predicted and filtered estimates are

x^t∣t−1=E[xt∣It−1],(4)\hat{x}_{t|t-1} = \mathbb{E}\left[x_t \mid\mathcal{I}_{t-1}\right], \tag*{(4)}
x^t∣t=E[xt∣It].(5)\hat{x}_{t|t} = \mathbb{E}\left[x_t \mid\mathcal{I}_t\right]. \tag*{(5)}

Here, It\mathcal{I}_t includes the initial information, received data, packet-reception history, and applied control inputs through time tt. When δt=0\delta_t = 0, no measurement update is performed, so x^t∣t=x^t∣t−1\hat{x}_{t|t} = \hat{x}_{t|t-1}.

For the nominal LQG problem with packet losses, the separation principle holds as studied in [41]. Therefore, we use the linear feedback law ut=−Lx^t∣tu_t = -L\hat{x}_{t|t}. The control input is transmitted over the lossless controller–actuator link.

Eavesdropping model

We consider an eavesdropper that can intercept packets transmitted over both the sensor–estimator and controller–actuator links. The eavesdropper knows the system model and uses its own estimator to reconstruct the system state from the intercepted signals. The sensor–eavesdropper link is assumed to be lossy. Its packet-loss process is independent of that of the legitimate sensor–estimator link, and the two links may have different packet-arrival probabilities.

The defence mechanism aims to prevent accurate state reconstruction by the eavesdropper despite its knowledge of the system model and access to intercepted packets. We evaluate its effectiveness using the eavesdropper’s state-estimation error and the expected time until the eavesdropper loses track of the system state.

Attack model

We consider a measurement replacement attack similar to that described in [23, 42]. The attacker aims to mislead the estimator by replacing legitimate sensor packets with fabricated packets generated by a stochastic dynamical model. The attack begins at an unknown time tat_a and persists indefinitely.

For every t≥tat \ge t_a, the transmitted sensor packet is replaced by zˉt=zta\bar{z}_t = z_t^a, where

zt+1a=Aazta+wta,wta∼i.i.d.N(0,Qa),(6)z_{t+1}^a = A_a z_t^a + w_t^a,\qquad w_t^a \mathrel{\overset{\mathrm{i.i.d.}}{\sim}} \mathcal{N}(0,Q_a), \tag*{(6)}

with Qa≻0Q_a \succ0 and ρ(Aa)<1\rho(A_a) < 1. The attack process is initialized in its stationary distribution,

ztaa∼N(0,Σz),z_{t_a}^a \sim\mathcal{N}(0,\Sigma_z),

where Σz≻0\Sigma_z \succ0 is the unique solution to

Σz=AaΣzAa⊤+Qa.\Sigma_z = A_a\Sigma_z A_a^{\top} + Q_a.

We assume, the initial attack state is independent of the attack noise sequence {wta}t≥ta\{w_t^a\}_{t\ge t_a}. The entire attack process is independent of x0x_0, ww, vv, ee, and δ\delta. Fabricated packets are subject to the same packet-loss mechanism as genuine packets.

Proposed Defence Mechanisms

In this section, we propose defence mechanisms to protect the NCS against the measurement replacement attacks and ensure privacy of the internal states of the system. First, we introduce an encoding-decoding scheme that helps maintain the confidentiality of the system states in the presence of eavesdroppers but not degrade the performance of the legitimate estimator.

Defence Against Eavesdropping: Observation Encoding-Decoding Scheme

The objective of the observation encoding-decoding scheme is to ensure that the legitimate estimator can accurately reconstruct the system states while preventing eavesdroppers from gaining useful information about the internal states of the system.

Let τt:=max⁡{s<t:δs=1}\tau_t := \max\{s < t : \delta_s = 1\} be the last reception time before tt. The sensor encodes and transmits the following signal to the legitimate estimator:

zˉt=yt−Mktyτt,where kt=t−τt.(7)\bar{z}_t = y_t - M^{k_t}y_{\tau_t},\qquad\text{where } k_t = t-\tau_t. \tag*{(7)}

Here M∈Rp×pM \in\mathbb{R}^{p\times p} is a matrix with at least one eigenvalue strictly greater than 1 and is known to both the encoder and the legitimate estimator.

Let δtu∈{0,1}\delta_t^u \in\{0,1\} indicate whether the packet zˉt\bar{z}_t is received by the legitimate estimator. The decoder then reconstructs

ytd=δtu(zˉt+Mktyτtd).(8)y_t^d = \delta_t^u\left(\bar{z}_t + M^{k_t}y_{\tau_t}^d\right). \tag*{(8)}

For the legitimate estimator, this ensures that it can accurately reconstruct the system states as long as it receives the encoded packets. However, for an eavesdropper there could be a critical event say at time tct_c with it missing a packet but the legitimate estimator receiving it. Once this happens, the eavesdropper loses the reference yτtcy_{\tau_{t_c}} and can no longer reconstruct the subsequent measurements. Furthermore, note that without attack, zˉt=zt\bar{z}_t = z_t. As shown in Lemma 1, the state estimation error of the eavesdropper grows unbounded after such a critical event, ensuring the confidentiality of the system states. On the other hand, the legitimate estimator continues to accurately reconstruct the system states as long as it receives the encoded packets.

Defence Against Measurement Replacement Attacks

To protect the NCS against measurement replacement attacks, we propose a mechanism that allows the legitimate estimator to detect such attacks as early as possible.

First, similar to several existing approaches, we have added a physical watermarking signal et∼N(0,Σe)e_t \sim\mathcal{N}(0,\Sigma_e) to the transmitted measurements. Specifically, the controller transmits

ut+et.(9)u_t + e_t. \tag*{(9)}

In our study, we take ete_t to be a random signal known to the legitimate estimator but unknown to the attacker. As studied before [23, 36], the addition of such a watermarking signal increases the detectability of data deception attacks.

Test Signal. We have used a combination of the innovation signal (see (A.10)) from the state estimator and the watermarking signal as the test signal to detect measurement replacement attacks. The state estimation equation with the watermarking signal and lossy measurements are given as

x^t∣t−1=(A+BL)x^t−1∣t−1+Bet−1,(10)\hat{x}_{t|t-1}=(\mathbf{A}+\mathbf{B}\mathbf{L})\hat{x}_{t-1|t-1}+\mathbf{B}e_{t-1}, \tag*{(10)}
Pt∣t−1=APt−1∣t−1AT+Q,(11)\mathbf{P}_{t|t-1}=\mathbf{A}\mathbf{P}_{t-1|t-1}\mathbf{A}^{\mathsf{T}}+\mathbf{Q}, \tag*{(11)}
Γt=CPt∣t−1CT+R,Kt=Pt∣t−1CTΓt−1,(12)\boldsymbol{\Gamma}_{t}=\mathbf{C}\mathbf{P}_{t|t-1}\mathbf{C}^{\mathsf{T}}+\mathbf{R},\qquad\mathbf{K}_{t}=\mathbf{P}_{t|t-1}\mathbf{C}^{\mathsf{T}}\boldsymbol{\Gamma}_{t}^{-1}, \tag*{(12)}
x^t∣t=x^t∣t−1+δtuKtγt,(13)\hat{x}_{t|t}=\hat{x}_{t|t-1}+\delta_{t}^{u}\mathbf{K}_{t}\gamma_{t}, \tag*{(13)}
Pt∣t=Pt∣t−1−δtuKtCPt∣t−1.(14)\mathbf{P}_{t|t}=\mathbf{P}_{t|t-1}-\delta_{t}^{u}\mathbf{K}_{t}\mathbf{C}\mathbf{P}_{t|t-1}. \tag*{(14)}

Finally, the innovation signal is given as provided the packet arrives,

γt=ytd−Cx^t∣t−1.(15)\gamma_{t}=y_{t}^{d}-\mathbf{C}\hat{x}_{t|t-1}. \tag*{(15)}

Here Pt∣t−1=E[(xt−x^t∣t−1)(xt−x^t∣t−1)T∣It−1]\mathbf{P}_{t|t-1}=\mathbb{E}[(x_{t}-\hat{x}_{t|t-1})(x_{t}-\hat{x}_{t|t-1})^{\mathsf{T}}\mid\mathcal{I}_{t-1}] is the predicted error covariance matrix, and Pt∣t=E[(xt−x^t∣t)(xt−x^t∣t)T∣It]\mathbf{P}_{t|t}=\mathbb{E}[(x_{t}-\hat{x}_{t|t})(x_{t}-\hat{x}_{t|t})^{\mathsf{T}}\mid\mathcal{I}_{t}] is the updated error covariance matrix.

Test statistic. We use a CUSUM test statistic [35] to detect the presence of a measurement replacement attack. The statistic recursively accumulates the log-likelihood ratio of the joint innovation and watermark signals under the attack hypothesis relative to the nominal hypothesis. The recursion is given in Theorem 1.

Theorem 1: Let Assumption 1 hold. The CUSUM test statistic to detect the measurement replacement attack (6) under packet dropouts and the encoding–decoding scheme (7)–(8) takes the following form:

St=max⁡{0,Wt},W0=0,(16)S_{t}=\max\{0,W_{t}\},\qquad W_{0}=0, \tag*{(16)}
Wt={max⁡{Wt−1+ℓtg,Wt−1+ℓtf},δtu=1,Wt−1,δtu=0,(17)W_{t}= \begin{cases} \max\{W_{t-1}+\ell_{t}^{g},W_{t-1}+\ell_{t}^{f}\}, & \delta_{t}^{u}=1,\\ W_{t-1}, & \delta_{t}^{u}=0, \end{cases} \tag*{(17)}

where, for j∈{g,f}j\in\{g,f\},

ℓtj=ln⁡fγ~tj(γ~t∣It−1)fγt(γ~t∣It−1).(18)\ell_{t}^{j}=\ln\frac{f_{\tilde{\gamma}_{t}}^{j}(\tilde{\gamma}_{t}\mid\mathcal{I}_{t-1})}{f_{\gamma_{t}}(\tilde{\gamma}_{t}\mid\mathcal{I}_{t-1})}. \tag*{(18)}

Here fγ~tff_{\tilde{\gamma}_{t}}^{f} and fγ~tgf_{\tilde{\gamma}_{t}}^{g} are the densities of the attacked innovation when the last received packet zˉτt\bar{z}_{\tau_{t}} is falsified (ta≤τtt_{a}\leq\tau_{t}) or genuine (τt<ta\tau_{t}<t_{a}), respectively,

{γ~t∣It−1,δtu=1}∼N(μtj,Σtj),(19)\{\tilde{\gamma}_{t}\mid\mathcal{I}_{t-1},\delta_{t}^{u}=1\}\sim\mathcal{N}(\mu_{t}^{j},\Sigma_{t}^{j}), \tag*{(19)}
μtj=Mktyτtd−C(A+BL)x^t−1∣t−1−CBet−1+{Kaktzˉτt,j=f (ta≤τt),0,j=g (τt<ta),(20)\mu_{t}^{j}=\mathbf{M}^{k_{t}}y_{\tau_{t}}^{d}-\mathbf{C}(\mathbf{A}+\mathbf{B}\mathbf{L})\hat{x}_{t-1|t-1}-\mathbf{C}\mathbf{B}e_{t-1}+ \begin{cases} \mathbf{K}_{a}^{k_{t}}\bar{z}_{\tau_{t}}, & j=f\ (t_{a}\leq\tau_{t}),\\ 0, & j=g\ (\tau_{t}<t_{a}), \end{cases} \tag*{(20)}
Σtj={Λkt:=∑i=0kt−1AaiQa(Aai)T,j=f,Σz,j=g,(21)\Sigma_{t}^{j}= \begin{cases} \boldsymbol{\Lambda}_{k_{t}}:=\sum_{i=0}^{k_{t}-1}\mathbf{A}_{a}^{i}\mathbf{Q}_{a}(\mathbf{A}_{a}^{i})^{\mathsf{T}}, & j=f,\\ \boldsymbol{\Sigma}_{z}, & j=g, \end{cases} \tag*{(21)}

and, in the absence of an attack,

{γt∣It−1,δtu=1}∼N(0,Γt),(22)\{\gamma_{t}\mid\mathcal{I}_{t-1},\delta_{t}^{u}=1\}\sim\mathcal{N}(0,\boldsymbol{\Gamma}_{t}), \tag*{(22)}

where Γt\boldsymbol{\Gamma}_{t} given in (12).

Proof: The proof of Theorem 1 is provided in Appendix VII-A.

Hypothesis testing framework

In this subsection, we formulate the detection of the measurement replacement attack as a hypothesis testing problem. At each time step tt, we consider the following hypotheses:

  • H0\mathcal{H}_{0}: No attack is present, i.e., the received innovation γt\gamma_{t} follows the nominal distribution (22).

  • H1\mathcal{H}_{1}: An attack is present, i.e., the received innovation γ~t\tilde{\gamma}_{t} follows the attacked distribution (19).

We perform the following test based on the test statistic StS_{t} to decide between the two hypotheses:

Decide H1 if St≥h, otherwise decide H0,(23)\text{Decide }\mathcal{H}_{1}\text{ if }S_{t}\geq h,\text{ otherwise decide }\mathcal{H}_{0}, \tag*{(23)}

where h>0h>0 is a predefined threshold. Under the CUSUM test, which requires a set of conditions to be satisfied (see [35]), the optimal threshold is log⁡(ARLh)\log(ARL_{h}), where ARLhARL_{h} denotes the user defined threshold on the average run length to false alarm (ARL), see (24). In practice, ARLhARL_{h} is chosen based on the desired trade-off between detection speed and false alarm rate.

ARL=E[inf⁡{t≥1:St≥h}∣H0],(24)ARL=\mathbb{E}[\inf\{t\geq1:S_{t}\geq h\}\mid\mathcal{H}_{0}], \tag*{(24)}

which represents the expected time to false alarm under the null hypothesis.

However, the conditions for optimality of the CUSUM may not be satisfied for the current detection problem, and the chosen threshold hh may only provide a suboptimal trade-off between detection speed and false alarm rate.

In practice, the attacker’s model may be unknown. In such cases, a joint estimation and detection scheme can estimate the post-change distributions online while detecting attacks, as demonstrated in [23]. The present study assumes a known attacker model to benchmark the performance of CUSUM-based detection, which is also the best-case scenario, and to investigate the fundamental trade-offs between confidentiality and attack detectability. Future work could extend the current framework to accommodate unknown or time-varying attacker models.

Statistical Performance Analysis

In this section, we analyse the statistical performance of the proposed defence mechanisms, focusing on the trade-off between the defence against eavesdropping and attacks. We derive analytical expressions for key performance metrics, such as the average time before the critical event, estimation error covariance for the eavesdropper and the legitimate receiver, the worst-case expected KLD, and other relevant statistical measures.

Performance Metrics for defence against eavesdropping

Our objective is to achieve perfect secrecy as defined in Definition 1. We adopt the notion of perfect secrecy of [13], adapted to the transmission of encoded measurements.

Definition 1 (Perfect secrecy): An encoding–decoding scheme (7)–(8) achieves perfect secrecy if and only if both of the following conditions hold:

(i) the legitimate user’s performance is optimal, i.e., for every tt,

x^t∣tu=E[xt∣It],Pt∣tu=Cov⁡(xt∣It),\hat{x}^{u}_{t|t}=\mathbb{E}[x_t\mid\mathcal{I}_t],\qquad P^{u}_{t|t}=\operatorname{Cov}(x_t\mid\mathcal{I}_t),

where It\mathcal{I}_t is the information set in (5). In other words, the encoding does not degrade the user’s estimate.

(ii) the eavesdropper’s error diverges to infinity with probability one,

tr⁡(Pt∣te)→a.s.∞as t→∞.\operatorname{tr}(P^{e}_{t|t})\xrightarrow[\text{a.s.}]{}\infty\quad\text{as }t\to\infty.

Next we introduce the key assumptions and metrics used to evaluate the performance of the defence mechanisms against eavesdropping.

Assumption 2 (Reception processes): Each packet is transmitted with a constant energy E>0E>0. The reception indicators δtu,δte∈{0,1}\delta^{u}_t,\delta^{e}_t\in\{0,1\} at the user and at the eavesdropper satisfy:

  1. the pairs {(δtu,δte)}t≥1\{(\delta^{u}_t,\delta^{e}_t)\}_{t\geq1} are i.i.d. over tt;

  2. for each tt, δtu\delta^{u}_t and δte\delta^{e}_t are independent, with

P{δtu=1}=δˉu,P{δte=1}=δˉe.\mathbb{P}\{\delta^{u}_t=1\}=\bar{\delta}_u,\qquad\mathbb{P}\{\delta^{e}_t=1\}=\bar{\delta}_e.
  1. we assume δˉu\bar{\delta}_u and δˉe\bar{\delta}_e always >δˉc>\bar{\delta}_c, where δˉc=max⁡(0.5,δ~c)\bar{\delta}_c=\max(0.5,\tilde{\delta}_c), where δ~c\tilde{\delta}_c is the critical arrival probability required for system’s stability [41].

Theorem 2 (Average time before the critical event): Let Assumption 2 hold with δˉu∈(0,1)\bar{\delta}_u\in(0,1) and δˉe∈(0,1)\bar{\delta}_e\in(0,1). Define the critical event and its first occurrence time

Ct:={δtu=1,δte=0},tc:=min⁡{t≥1:Ct occurs},C_t:=\{\delta^{u}_t=1,\delta^{e}_t=0\},\qquad t_c:=\min\{t\geq1:C_t\text{ occurs}\},

and let pc:=δˉu(1−δˉe)∈(0,1)p_c:=\bar{\delta}_u(1-\bar{\delta}_e)\in(0,1). The average time before the critical event (ATBCE) is

E[tc]=1δˉu(1−δˉe).(25)\mathbb{E}[t_c]=\frac{1}{\bar{\delta}_u(1-\bar{\delta}_e)}. \tag*{(25)}

Assuming, δˉu=δˉe\bar{\delta}_u=\bar{\delta}_e, the ATBCE relates to the transmission energy EE as

E[tc]=1ϵE(1−ϵE)\mathbb{E}[t_c]=\frac{1}{\epsilon^{E}(1-\epsilon^{E})}

where ϵ\epsilon is given in (3).

Proof. The proof of Theorem 2 is provided in Appendix VII-B.

Assumption 3 (Eavesdropper): The eavesdropper knows (A,B,C,Q,R,M)(A,B,C,Q,R,M) and overhears the acknowledgements and the applied input ut+etu_t+e_t; hence it knows τt\tau_t and tct_c. The encoder, the user and the eavesdropper share the initial reference y0y_0, i.e., τt:=0\tau_t:=0 if δsu=0\delta^{u}_s=0 for all 1≤s<t1\leq s<t, and r0e:=y0r^e_0:=y_0. At every reception (δte=1\delta^{e}_t=1) the eavesdropper applies the legitimate decoder to its own references,

yte=zt+Mktrte,(26)y^e_t=z_t+M^{k_t}r^e_t, \tag*{(26)}

and feeds ytey^e_t to the Kalman filter (10)–(14) with δt\delta_t replaced by δte\delta^{e}_t, producing x^t∣t−1e,x^t∣te\hat{x}^{e}_{t|t-1},\hat{x}^{e}_{t|t}. At every tt with δtu=1\delta^{u}_t=1 it stores the reference

rte={yte,δte=1,Cx^t∣t−1e,δte=0.r^e_t= \begin{cases} y^e_t, & \delta^{e}_t=1,\\ C\hat{x}^{e}_{t|t-1}, & \delta^{e}_t=0. \end{cases}

Lemma 1: Let ρ(M)>1\rho(M)>1 and let Assumption 1–3 hold. Define the state estimation error covariances of the legitimate user and the eavesdropper as PuP^u and PeP^e. Then the following statements hold.

(i) Legitimate user: The legitimate user’s state estimation will satisfy the perfect secrecy condition as stated in Definition 1(i).

(ii) Eavesdropper: Suppose the critical event, as defined in Theorem 2, occurs at tct_c and that δte=1\delta^{e}_t=1 for all t>tct>t_c, which is the best case scenario for the eavesdropper. Let (A,C)(A,C) be detectable, (A,Q1/2)(A,Q^{1/2}) stabilizable, and CQCT≻0CQC^{\mathsf T}\succ0, and let Pt∣te:=E[x~t∣tex~t∣teT]P^e_{t|t}:=\mathbb{E}[\tilde{x}^{e}_{t|t}\tilde{x}^{e\mathsf T}_{t|t}], with x~t∣te:=xt−x^t∣te\tilde{x}^{e}_{t|t}:=x_t-\hat{x}^{e}_{t|t}, be the actual error covariance of the eavesdropper’s estimate. Then there exist a constant c>0c>0, independent of tct_c, and a finite T0T_0 such that

tr⁡(Pt∣te)≥cρ(M)2(t−tc),t≥tc+T0.\operatorname{tr}(P^e_{t|t})\geq c\rho(M)^{2(t-t_c)},\qquad t\geq t_c+T_0.

Hence tr⁡(Pt∣te)→∞\operatorname{tr}(P^e_{t|t})\to\infty as $t\to\infty.

Proof. Part (i) follows from [41], since the user decodes every received packet exactly. Part (ii) is proved in Appendix VII-C.

Performance Metrics for defence against attacks:

Here our objective is to detect the attack as soon as possible after it occurs. In other words, we would like to have a KLD between the post-attack and pre-attack distributions that is as large as possible. Under the classical i.i.d. assumptions, the worst-case average detection delay (WADD) of the optimal CUSUM test scales asymptotically as the detection threshold divided by the KLD between the post- and pre-attack distributions [35]. Applying this relationship to our setting, however, requires regularity conditions that have not been established. Nevertheless, KLD remains a useful measure of the statistical distinguishability of nominal and attacked signals. In our setting, packet losses make the state-estimation error covariance depend on the reception history, rather than converge to a deterministic steady-state value. Consequently, the conditional KLD also depends on this history, making its expectation difficult to evaluate analytically. We therefore use the worst-case expected KLD as a tractable, conservative measure of attack detectability, without asserting a corresponding bound on WADD.

Remark 1: Under the persistent measurement replacement attack, the first packet received at or after tat_a is decoded using a genuine reference (τt<ta\tau_t<t_a, j=gj=g). All subsequent receptions use a reference obtained from a falsified packet (τt≥ta\tau_t\geq t_a, j=fj=f). The following KLD analysis therefore focuses on j=fj=f, which applies after the first attacked reception. The detector still evaluates both likelihoods because the attack onset is unknown.

Theorem 3 (Worst-case expected KLD): Let Assumption 1–3 hold and (A,C)(A,C) be detectable and (A,Q1/2)(A,Q^{1/2}) stabilizable. Suppose that the last received packet is falsified, i.e., ta≤τtt_a\leq\tau_t (j=fj=f). Then the expected conditional KLD between the post- and pre-attack conditional distribution used in the CUSUM test in Theorem 1, with fγtff^f_{\gamma_t} and fytff^f_{y_t} given in (19) and (22), and

with the convention that the conditional KLD is zero when δtu=0\delta_t^u=0 (no packet is received and the CUSUM statistic (17) is not updated), satisfies in steady state

E[D(fy~tff,fγt∣It−1)]≥WKLDf,(27)\mathbb{E}\left[D\left(f_{\tilde{y}_t^f}^{f},f_{\gamma_t}\mid\mathcal{I}_{t-1}\right)\right]\geq\mathrm{WKLD}_f, \tag*{(27)}

where the worst-case expected KLD, WKLDf\mathrm{WKLD}_f, is given by

WKLDf=δˉu2(tr⁡((CPˉC⊤+R)−1(Qa+N))−p+log⁡∣CP⋆C⊤+R∣∣Σz∣),(28)\mathrm{WKLD}_f=\frac{\bar{\delta}^u}{2}\left(\operatorname{tr}\left((C\bar{P}C^{\top}+R)^{-1}(Q_a+N)\right)-p+\log\frac{\lvert CP^{\star}C^{\top}+R\rvert}{\lvert\Sigma_z\rvert}\right), \tag*{(28)}

where N:=CBΣeB⊤C⊤⪰0N:=CB\Sigma_eB^{\top}C^{\top}\succeq0. Here, Pˉ\bar{P} is the stabilizing solution of the MARE

Pˉ=APˉA⊤+Q−δˉuAPˉC⊤(CPˉC⊤+R)−1CPˉA⊤,\bar{P}=A\bar{P}A^{\top}+Q-\bar{\delta}^uA\bar{P}C^{\top}(C\bar{P}C^{\top}+R)^{-1}C\bar{P}A^{\top},

which satisfies lim sup⁡t→∞E[Pt∣t−1]⪯Pˉ\limsup_{t\to\infty}\mathbb{E}[P_{t\mid t-1}]\preceq\bar{P} [41], and P⋆P^{\star} is the stabilizing solution of the standard ARE (the MARE with δˉu=1\bar{\delta}^u=1)

P⋆=AP⋆A⊤+Q−AP⋆C⊤(CP⋆C⊤+R)−1CP⋆A⊤,P^{\star}=AP^{\star}A^{\top}+Q-AP^{\star}C^{\top}(CP^{\star}C^{\top}+R)^{-1}CP^{\star}A^{\top},

which is the error covariance of the lossless Kalman filter and lower bounds Pt∣t−1P_{t\mid t-1} for every packet-reception realization, i.e., P⋆⪯Pt∣t−1P^{\star}\preceq P_{t\mid t-1} almost surely in steady state.

Proof: The proof of Theorem 3 is provided in Appendix VII-D.

Remark 2: The conditional KLD D(fy~tjj,fγt∣It−1)D\left(f_{\tilde{y}_t^j}^{j},f_{\gamma_t}\mid\mathcal{I}_{t-1}\right) is a random variable, since Γt\Gamma_t, μtj\mu_t^j and Σtj\Sigma_t^j depend on the random packet-reception scenario. As a result, its expected value has no closed-form expression. The WKLD in Theorem 3 overcomes this by providing a closed-form bound that depends only on the system matrices and on the parameters δˉ\bar{\delta} and Σe\Sigma_e, and can therefore be used to select the transmission energy and the watermark.

Remark 3 (Dependence on transmission energy): By (3), δˉu=1−ϵE\bar{\delta}^u=1-\epsilon^E increases with EE, and Pˉ\bar{P} is nonincreasing in δˉu\bar{\delta}^u [41]. Hence the trace term of (28), tr⁡((CPˉC⊤+R)−1(Qa+N))\operatorname{tr}((C\bar{P}C^{\top}+R)^{-1}(Q_a+N)), which carries the watermark contribution NN, is nondecreasing in EE. A higher transmission energy thus makes the watermark more visible in the innovation. The log-determinant term of (28) depends only on P⋆P^{\star} and Σz\Sigma_z and is therefore independent of EE. Consequently, the bracketed term in (28) is nondecreasing in EE, and whenever it is positive, i.e., whenever the bound is informative, WKLDf\mathrm{WKLD}_f itself is nondecreasing in EE through both the prefactor δˉu\bar{\delta}^u and the trace term. The resulting trade-off with the ATBCE of Theorem 2 is evaluated numerically in Section V.

NUMERICAL RESULTS

In this section, we evaluate the proposed defence mechanisms through Monte Carlo simulations. We examine the effect of observation encoding on state estimation, assess the performance of the watermarking-based CUSUM detector, and compare the expected KLD with its analytical bound. We then study how transmission energy influences confidentiality and attack detection.

Figure 2. Trace of the empirical and Kalman-predicted error covariances for the user (uu) and the eavesdropper (ee), averaged over NN Monte Carlo runs. The vertical line marks the switching time kc=6k_c=6.

Figure 3. Evolution of the CUSUM statistic SkS_k. The attack starts at tat_a, and an alarm is raised at kdk_d, the first time SkS_k exceeds the threshold h=log⁡(ARL0)h=\log(\mathrm{ARL}_0) with ARL0=106\mathrm{ARL}_0=10^6.

Simulation Setup

We consider the linear stochastic system in (1)–(2), with the parameters listed in Appendix ??. The legitimate receiver uses the estimator and controller described in Sections II and III. Observations are encoded according to (7), with ρ(M)>1\rho(M)>1, and the controller adds an independent Gaussian watermark with covariance Σe\Sigma_e to the applied input.

Packet reception follows (3). The legitimate and eavesdropper reception processes are independent and have equal arrival probabilities, δˉu=δˉe\bar{\delta}_u=\bar{\delta}_e, unless otherwise stated. We vary the transmission energy while keeping the remaining channel parameters fixed. The operating range used for estimation and detection comparisons satisfies δˉu>δˉe\bar{\delta}_u>\bar{\delta}_e.

The replacement attack begins at tat_a and follows (6), with ρ(Aa)<1\rho(A_a)<1 and stationary initialization. Fabricated packets experience the same losses as genuine packets. Results are averaged over NMCN_{\mathrm{MC}} independent trials, with [confidence level] confidence intervals. Comparisons between schemes use matched noise and reception realizations.

CONCLUSIONS

We study joint protection of an NCS against eavesdropping and measurement-replacement attacks over lossy wireless channels. The defence mechanism preserves the legitimate receiver’s estimation performance during nominal operation, while the eavesdropper’s state-estimation error covariance grows without bound after a critical event. We characterised the relationship between the expected time to this critical event and the transmission energy. Higher transmission energy increases the expected time to the critical event, meaning the eavesdropper can reliably reconstruct the system state longer before the critical event occurs. On the other hand, defence against measurement replacement attacks relies on timely detection, characterised by the KLD between the post- and pre-attack distributions of the test signal. However, in the lossy wireless-channel scenario, the conditional KLD depends on the reception history, so it does not admit a simple closed-form expression. Therefore, we derive the worst-case expected KLD as a conservative measure of attack detectability. The WKLD provides a lower bound on expected attack detectability, guiding the design of transmission energy and watermarking strategies to balance confidentiality and timely attack detection.

APPENDIX

Proof of Theorem 1

Pre-attack. Let It={γ1:t,e1:t}\mathcal{I}_t = \{\gamma_{1:t}, e_{1:t}\} denote the information available at time tt, and recall that the innovation is γt=ytd−Cx^t∣t−1\gamma_t = y_t^d - C\hat{x}_{t|t-1}. If δtu=1\delta_t^u = 1, then

γt=yt−Cx^t∣t−1=Cxt+vt−Cx^t∣t−1=C(Axt−1+B(Lx^t−1∣t−1+et−1)+wt−1)+vt−C(Ax^t−1∣t−1+B(Lx^t−1∣t−1+et−1))=CA(xt−1−x^t−1∣t−1)+Cwt−1+vt.(29)\begin{aligned} \gamma_t &= y_t - C\hat{x}_{t|t-1} = Cx_t + v_t - C\hat{x}_{t|t-1}\\ &= C\left(Ax_{t-1} + B(L\hat{x}_{t-1|t-1} + e_{t-1}) + w_{t-1}\right) + v_t\\ &\quad- C\left(A\hat{x}_{t-1|t-1} + B(L\hat{x}_{t-1|t-1} + e_{t-1})\right)\\ &= CA(x_{t-1} - \hat{x}_{t-1|t-1}) + Cw_{t-1} + v_t. \tag*{(29)} \end{aligned}

Since the estimation error, wt−1w_{t-1}, and vtv_t are zero mean, E[γt]=0\mathbb{E}[\gamma_t] = 0. If δtu=0\delta_t^u = 0, then γt=0\gamma_t = 0. Hence, E[δtuγt]=0\mathbb{E}[\delta_t^u\gamma_t] = 0, and, since δtu\delta_t^u is independent of γt\gamma_t,

cov⁡(δtuγt)=E[(δtu)2γtγt⊤]=E[(δtu)2]E[γtγt⊤]=δˉu(CPt∣t−1C⊤+R).(30)\begin{aligned} \operatorname{cov}(\delta_t^u\gamma_t) &= \mathbb{E}\left[(\delta_t^u)^2\gamma_t\gamma_t^\top\right] = \mathbb{E}\left[(\delta_t^u)^2\right]\mathbb{E}\left[\gamma_t\gamma_t^\top\right]\\ &= \bar{\delta}^u\left(CP_{t|t-1}C^\top+ R\right). \tag*{(30)} \end{aligned}

We next show that the innovation sequence is white. By the tower property,

E[γt∣γ1:t−1,e1:t−1]=E[E[γt∣It−1]∣γ1:t−1,e1:t−1],(31)\mathbb{E}[\gamma_t \mid\gamma_{1:t-1},e_{1:t-1}] = \mathbb{E}\left[\mathbb{E}[\gamma_t \mid\mathcal{I}_{t-1}] \mid\gamma_{1:t-1},e_{1:t-1}\right], \tag*{(31)}

where

E[γt∣It−1]=δˉuE[xt−x^t∣t−1∣It−1]=δˉuE[xt−E[xt∣It−1]∣It−1]=0.(32)\begin{aligned} \mathbb{E}[\gamma_t \mid\mathcal{I}_{t-1}] &= \bar{\delta}^u\mathbb{E}[x_t-\hat{x}_{t|t-1} \mid\mathcal{I}_{t-1}]\\ &= \bar{\delta}^u\mathbb{E}[x_t-\mathbb{E}[x_t \mid\mathcal{I}_{t-1}] \mid\mathcal{I}_{t-1}]\\ &= 0. \tag*{(32)} \end{aligned}

Therefore, E[γt∣γ1:t−1,e1:t−1]=0\mathbb{E}[\gamma_t \mid\gamma_{1:t-1},e_{1:t-1}] = 0, i.e., γt\gamma_t remains white. Finally, by the law of total covariance,

cov⁡(γt∣It−1)=E[cov⁡(γt∣It−1,δtu)]+cov⁡(E[γt∣It−1,δtu]).(33)\operatorname{cov}(\gamma_t \mid\mathcal{I}_{t-1}) = \mathbb{E}\left[\operatorname{cov}(\gamma_t \mid\mathcal{I}_{t-1},\delta_t^u)\right] + \operatorname{cov}\left(\mathbb{E}[\gamma_t \mid\mathcal{I}_{t-1},\delta_t^u]\right). \tag*{(33)}

APPENDIX

Conditioning on δtu=1\delta_t^u = 1 yields

cov⁡(γt∣It−1,δtu=1)=CPt∣t−1C⊤+R=Γt,(34)\operatorname{cov}(\gamma_t \mid\mathcal{I}_{t-1},\delta_t^u = 1) = CP_{t|t-1}C^\top+ R = \Gamma_t, \tag*{(34)}

which establishes (22). Under attack. Let t=tat=t_a denote the time instant at which the attack starts. Under attack, the fake measurements evolve according to

zta=Aazt−1a+wa,t−1,(35)z_t^a = A_a z_{t-1}^a + w_{a,t-1}, \tag*{(35)}

and zˉt=zta\bar{z}_t = z_t^a after the attack. The received signal is

yˉt={zˉt,δtu=1,0,δtu=0,(36)\bar{y}_t = \begin{cases} \bar{z}_t, & \delta_t^u = 1,\\ 0, & \delta_t^u = 0, \end{cases} \tag*{(36)}

and the decoded measurements are

ytd={zˉt+Mt−tkytk,δtu=1,0,δtu=0.(37)y_t^d = \begin{cases} \bar{z}_t + M^{t-t_k}y_{t_k}, & \delta_t^u = 1,\\ 0, & \delta_t^u = 0. \end{cases} \tag*{(37)}

Two cases must be distinguished, depending on the position of tkt_k relative to the attack start tat_a. In both cases, the innovation signal is

γt=ytd−Cx^t∣t−1.(38)\gamma_t = y_t^d - C\hat{x}_{t|t-1}. \tag*{(38)}

Case 1. We first consider Case 1, i.e., tk<tat_k < t_a and t≥tat \geq t_a. Then ytkd=ytky_{t_k}^d = y_{t_k} is the true observation, and

ytd={zta+Mt−tkytk,δtu=1,0,δtu=0.(39)y_t^d = \begin{cases} z_t^a + M^{t-t_k}y_{t_k}, & \delta_t^u = 1,\\ 0, & \delta_t^u = 0. \end{cases} \tag*{(39)}

The innovation becomes

γt=ytd−Cx^t∣t−1=zta+Mt−tkytk−C(A+BL)x^t−1∣t−1−CBet−1,=wa,t−1+Aazt−1a+Mt−tkytk−C(A+BL)x^t−1∣t−1−CBet−1,(40)\begin{aligned} \gamma_t &= y_t^d - C\hat{x}_{t|t-1}\\ &= z_t^a + M^{t-t_k}y_{t_k} - C(A+BL)\hat{x}_{t-1|t-1} - CBe_{t-1},\\ &= w_{a,t-1} + A_a z_{t-1}^a + M^{t-t_k}y_{t_k}\\ &\quad- C(A+BL)\hat{x}_{t-1|t-1} - CBe_{t-1}, \tag*{(40)} \end{aligned}

with conditional mean and covariance

E[γt∣Iγe]=Mt−tkytk−C(A+BL)x^t−1∣t−1−CBet−1,(41)\mathbb{E}[\gamma_t \mid\mathcal{I}_{\gamma_e}] = M^{t-t_k}y_{t_k} - C(A+BL)\hat{x}_{t-1|t-1} - CBe_{t-1}, \tag*{(41)}
cov⁡(γt∣Iγe)=E[ztazta⊤]=Σz.(42)\operatorname{cov}(\gamma_t \mid\mathcal{I}_{\gamma_e}) = \mathbb{E}[z_t^a z_t^{a\top}] = \Sigma_z. \tag*{(42)}

Indeed, since

γtk=ytk−Cx^tk∣tk−1,(43)\gamma_{t_k} = y_{t_k} - C\hat{x}_{t_k|t_k-1}, \tag*{(43)}

ytky_{t_k} is known or can be recoverable from Iγe\mathcal{I}_{\gamma_e}. Moreover, ta>tkt_a > t_k and t≥tat \geq t_a imply that no sample of zaz^a has been received up to time t−1t-1. Consequently, zt−1az_{t-1}^a is a random variable with zero mean and covariance Σz\Sigma_z, where Σz\Sigma_z is the solution of the Lyapunov equation

Σz=AaΣzAa⊤+Qa.(44)\Sigma_z = A_a\Sigma_z A_a^\top+ Q_a. \tag*{(44)}

Case 2. We now consider Case 2. For δtu=1\delta_t^u = 1,

γt=zta+Mt−tkytkd−Cx^t∣t−1=wa,t−1+Aazt−1a+Mt−tkytkd−C(Ax^t−1∣t−1+BLx^t−1∣t−1+Bet−1)=wa,t−1+Aazt−1a+Mt−tkytkd−C(A+BL)x^t−1∣t−1−CBet−1,(45)\begin{aligned} \gamma_t &= z_t^a + M^{t-t_k}y_{t_k}^d - C\hat{x}_{t|t-1}\\ &= w_{a,t-1} + A_a z_{t-1}^a + M^{t-t_k}y_{t_k}^d\\ &\quad- C(A\hat{x}_{t-1|t-1} + BL\hat{x}_{t-1|t-1} + Be_{t-1})\\ &= w_{a,t-1} + A_a z_{t-1}^a + M^{t-t_k}y_{t_k}^d\\ &\quad- C(A+BL)\hat{x}_{t-1|t-1} - CBe_{t-1}, \tag*{(45)} \end{aligned}

where

ytkd=δtku(γtk+C(A+BL)x^tk−1∣tk−1+CBetk−1)(46)y_{t_k}^{d}=\delta_{t_k}^{u}\left(\gamma_{t_k}+\mathrm{C}(\mathrm{A}+\mathrm{B}\mathrm{L})\hat{x}_{t_{k-1}|t_{k-1}}+\mathrm{C}\mathrm{B}e_{t_{k-1}}\right) \tag*{(46)}

and, for δt−1u=1\delta_{t-1}^{u}=1,

γt−1=yt−1d−Cx^t−1∣t−2=zt−1a+Mt−tkytkd−Cx^t−1∣t−2.(47)\begin{aligned} \gamma_{t-1}=y_{t-1}^{d}-\mathrm{C}\hat{x}_{t-1|t-2} \\ &=z_{t-1}^{a}+\mathrm{M}^{t-t_k}y_{t_k}^{d}-\mathrm{C}\hat{x}_{t-1|t-2}. \tag*{(47)} \end{aligned}

If δt−1u=0\delta_{t-1}^{u}=0, zt−1az_{t-1}^{a} cannot be reconstructed. However, δtku=1\delta_{t_k}^{u}=1, and iterating (A.7) gives

zta=Aat−tkztka+∑i=1t−tkAai−1wa,t−i.(48)z_{t}^{a}=\mathrm{A}_{a}^{t-t_k}z_{t_k}^{a}+\sum_{i=1}^{t-t_k}\mathrm{A}_{a}^{i-1}w_{a,t-i}. \tag*{(48)}

Define the information set Iγe={γ1t−1,e1t−1}\mathcal{I}_{\gamma_e}=\{\gamma_{1}^{t-1},e_{1}^{t-1}\}. Substituting (A.20) into (A.17) yields

γt=Aat−tkztka+∑i=1t−tkAai−1wa,t−i+Mt−tkytkd−C(A+BL)x^t−1∣t−1−CBet−1.(49)\gamma_t=\mathrm{A}_{a}^{t-t_k}z_{t_k}^{a}+\sum_{i=1}^{t-t_k}\mathrm{A}_{a}^{i-1}w_{a,t-i}+\mathrm{M}^{t-t_k}y_{t_k}^{d} -\mathrm{C}(\mathrm{A}+\mathrm{B}\mathrm{L})\hat{x}_{t-1|t-1}-\mathrm{C}\mathrm{B}e_{t-1}. \tag*{(49)}

Hence, the conditional mean is

E[γt∣Iγe]=E[γt∣x^t−1∣t−1,et−1,ytkd,γt−1],\mathbb{E}[\gamma_t\mid\mathcal{I}_{\gamma_e}] =\mathbb{E}\left[\gamma_t\mid\hat{x}_{t-1|t-1},e_{t-1},y_{t_k}^{d},\gamma_{t-1}\right],
μγ∣γe=Aat−tkztka+Mt−tkytkd−C(A+BL)x^t−1∣t−1−CBet−1=Aat−tkzˉtka+Mt−tkytkd−C(A+BL)x^t−1∣t−1−CBet−1.(50)\begin{aligned} \mu_{\gamma\mid\gamma_e} =\mathrm{A}_{a}^{t-t_k}z_{t_k}^{a}+\mathrm{M}^{t-t_k}y_{t_k}^{d} -\mathrm{C}(\mathrm{A}+\mathrm{B}\mathrm{L})\hat{x}_{t-1|t-1}-\mathrm{C}\mathrm{B}e_{t-1} \\ &=\mathrm{A}_{a}^{t-t_k}\bar{z}_{t_k}^{a}+\mathrm{M}^{t-t_k}y_{t_k}^{d} -\mathrm{C}(\mathrm{A}+\mathrm{B}\mathrm{L})\hat{x}_{t-1|t-1}-\mathrm{C}\mathrm{B}e_{t-1}. \tag*{(50)} \end{aligned}

For the conditional covariance,

cov⁡(γt∣Iγe)=E[(γt−E[γt∣Iγe])(γt−E[γt∣Iγe])T∣Iγe].(51)\operatorname{cov}(\gamma_t\mid\mathcal{I}_{\gamma_e}) =\mathbb{E}\left[(\gamma_t-\mathbb{E}[\gamma_t\mid\mathcal{I}_{\gamma_e}]) (\gamma_t-\mathbb{E}[\gamma_t\mid\mathcal{I}_{\gamma_e}])^{\mathsf{T}}\mid\mathcal{I}_{\gamma_e}\right]. \tag*{(51)}

where, from (A.21) and (A.22),

γt−E[γt∣Iγe]=∑i=1t−tkAai−1wa,t−i.(52)\gamma_t-\mathbb{E}[\gamma_t\mid\mathcal{I}_{\gamma_e}] =\sum_{i=1}^{t-t_k}\mathrm{A}_{a}^{i-1}w_{a,t-i}. \tag*{(52)}

Therefore,

cov⁡(γt∣Iγe)=E[(∑i=1t−tkAai−1wa,t−i)(∑i=1t−tkAai−1wa,t−i)T],\operatorname{cov}(\gamma_t\mid\mathcal{I}_{\gamma_e}) =\mathbb{E}\left[ \left(\sum_{i=1}^{t-t_k}\mathrm{A}_{a}^{i-1}w_{a,t-i}\right) \left(\sum_{i=1}^{t-t_k}\mathrm{A}_{a}^{i-1}w_{a,t-i}\right)^{\mathsf{T}} \right],
Σγ∣γe=∑i=1t−tkAai−1Qa(Aai−1)T.(53)\Sigma_{\gamma\mid\gamma_e} =\sum_{i=1}^{t-t_k}\mathrm{A}_{a}^{i-1}\mathrm{Q}_{a} \left(\mathrm{A}_{a}^{i-1}\right)^{\mathsf{T}}. \tag*{(53)}

Proof of Theorem 2

By Assumption 2, at each time instant t=1,2,…t=1,2,\ldots, the critical event CtC_t occurs with probability

P{Ct}=P{δtu=1}P{δte=0}=δˉu(1−δˉe)=pc,(54)\mathbb{P}\{C_t\}=\mathbb{P}\{\delta_t^u=1\}\mathbb{P}\{\delta_t^e=0\} =\bar{\delta}_u(1-\bar{\delta}_e)=p_c, \tag*{(54)}

independently across time. Since tc=tt_c=t if and only if C1,…,Ct−1C_1,\ldots,C_{t-1} do not occur and CtC_t occurs,

P{tc=t}=(1−pc)t−1pc,t=1,2,…,(55)\mathbb{P}\{t_c=t\}=(1-p_c)^{t-1}p_c,\qquad t=1,2,\ldots, \tag*{(55)}

that is, tct_c is geometrically distributed with success probability pc∈(0,1)p_c\in(0,1); in particular, tc<∞t_c<\infty almost surely. Consequently,

E[tc]=∑t=1∞t(1−pc)t−1pc=pc(1−(1−pc))2=1pc=1δˉu(1−δˉe).(56)\begin{aligned} \mathbb{E}[t_c]=\sum_{t=1}^{\infty}t(1-p_c)^{t-1}p_c =\frac{p_c}{(1-(1-p_c))^2} \\ &=\frac{1}{p_c}=\frac{1}{\bar{\delta}_u(1-\bar{\delta}_e)}. \tag*{(56)} \end{aligned}

where we used ∑t≥1txt−1=(1−x)−2\sum_{t\geq1}tx^{t-1}=(1-x)^{-2} for ∣x∣<1|x|<1. This proves (25).

Finally, let δˉu=δˉe=:δˉ\bar{\delta}_u=\bar{\delta}_e=:\bar{\delta}. By (3), δˉ=1−ϵE\bar{\delta}=1-\epsilon^E, and substituting into (B.3) gives

E[tc]=1δˉ(1−δˉ)=1ϵE(1−ϵE).(57)\mathbb{E}[t_c]=\frac{1}{\bar{\delta}(1-\bar{\delta})} =\frac{1}{\epsilon^E(1-\epsilon^E)}. \tag*{(57)}

Proof of Lemma 1(ii)

Let Πt∣t−1\Pi_{t|t-1} be the covariance computed by the eavesdropper’s Kalman filter, i.e., the Riccati recursion (11)–(14) with δt\delta_t replaced by δte\delta_t^e, let Kte=Πt∣t−1CT(CΠt∣t−1CT+R)−1K_t^e=\Pi_{t|t-1}\mathrm{C}^{\mathsf{T}}(\mathrm{C}\Pi_{t|t-1}\mathrm{C}^{\mathsf{T}}+\mathrm{R})^{-1}, Ft:=I−KteCF_t:=\mathrm{I}-K_t^e\mathrm{C}, and x~t∣t:=xt−x^t∣te\tilde{x}_{t|t}:=x_t-\hat{x}_{t|t}^e.

Step 1 (decoding mismatch). Let Δyt:=yt−yte\Delta y_t:=y_t-y_t^e and

η:=ytc−Cx^tc∣tc−1e=Cx~tc∣tc−1e+vtc.(58)\eta:=y_{t_c}-\mathrm{C}\hat{x}_{t_c|t_c-1}^e =\mathrm{C}\tilde{x}_{t_c|t_c-1}^e+v_{t_c}. \tag*{(58)}

For every user-reception time s≥tcs\geq t_c, let ϵs:=ys−rse\epsilon_s:=y_s-r_s^e. By Assumption 3, rtce=Cx^tc∣tc−1er_{t_c}^e=\mathrm{C}\hat{x}_{t_c|t_c-1}^e, so ϵtc=η\epsilon_{t_c}=\eta. For s>tcs>t_c with δsu=1\delta_s^u=1, we have δse=1\delta_s^e=1, so rse=yse=zs+Mksrτser_s^e=y_s^e=z_s+\mathrm{M}^{k_s}r_{\tau_s}^e, whereas (7) gives ys=zs+Mksyτsy_s=z_s+\mathrm{M}^{k_s}y_{\tau_s}. Hence ϵs=Mksϵτs\epsilon_s=\mathrm{M}^{k_s}\epsilon_{\tau_s}, and since ks=s−τsk_s=s-\tau_s, induction gives ϵs=Ms−tcη\epsilon_s=\mathrm{M}^{s-t_c}\eta. For any t>tct>t_c, τt≥tc\tau_t\geq t_c, and therefore

Δyt=Mktϵτt=Mt−tcη.(59)\Delta y_t=\mathrm{M}^{k_t}\epsilon_{\tau_t}=\mathrm{M}^{t-t_c}\eta. \tag*{(59)}

Step 2 (error dynamics). The eavesdropper knows the applied input ut−1+et−1u_{t-1}+e_{t-1}, so x~t∣t−1e=Ax~t−1∣t−1e+wt−1\tilde{x}_{t|t-1}^e=\mathrm{A}\tilde{x}_{t-1|t-1}^e+w_{t-1}. For t>tct>t_c, δte=1\delta_t^e=1 and yte=Cxt+vt−Δyty_t^e=\mathrm{C}x_t+v_t-\Delta y_t, so by (D.2)

x~t∣te=FtAx~t−1∣t−1e+Ftwt−1−Ktevt+KteMt−tcη.(60)\tilde{x}_{t|t}^e=F_t\mathrm{A}\tilde{x}_{t-1|t-1}^e+F_tw_{t-1}-K_t^ev_t+K_t^e\mathrm{M}^{t-t_c}\eta. \tag*{(60)}

Step 3 (lower bound). For t<tct<t_c, every user reception is also an eavesdropper reception, so the eavesdropper decodes exactly and its filter is matched. Hence x~tc∣tc−1e=Πtc∣tc−1\tilde{x}_{t_c|t_c-1}^e=\Pi_{t_c|t_c-1} and, since δtce=0\delta_{t_c}^e=0, x~tc∣tce=x~tc∣tc−1e\tilde{x}_{t_c|t_c}^e=\tilde{x}_{t_c|t_c-1}^e. As vtcv_{t_c} is independent of x~tc∣tc−1e\tilde{x}_{t_c|t_c-1}^e, (D.1) gives H:=E[ηηT]=CPtc∣tc−1CT+R⪰R≻0\mathrm{H}:=\mathbb{E}[\eta\eta^{\mathsf{T}}]=\mathrm{C}\mathrm{P}_{t_c|t_c-1}\mathrm{C}^{\mathsf{T}}+\mathrm{R}\succeq\mathrm{R}\succ0. Define Ψt:=E[x~t∣teηT]H−1\Psi_t:=\mathbb{E}[\tilde{x}_{t|t}^e\eta^{\mathsf{T}}]\mathrm{H}^{-1}. Since the second-moment matrix of (x~t∣te,η)(\tilde{x}_{t|t}^e,\eta) is positive semidefinite, its Schur complement gives Pt∣te⪰ΨtHΨtT⪰λmin⁡(R)ΨtΨtT\mathrm{P}_{t|t}^e\succeq\Psi_t\mathrm{H}\Psi_t^{\mathsf{T}}\succeq\lambda_{\min}(\mathrm{R})\Psi_t\Psi_t^{\mathsf{T}}. Let λ∈C\lambda\in\mathbb{C} be an eigenvalue of M\mathrm{M} with ∣λ∣=ρ(M)>1|\lambda|=\rho(\mathrm{M})>1, and let v\mathbf{v} be a corresponding eigenvector with ∥v∥=1\|\mathbf{v}\|=1. Then

tr⁡(Pt∣te)≥λmin⁡(R)∥Ψt∥F2≥λmin⁡(R)∥Ψtv∥2.(61)\operatorname{tr}(\mathrm{P}_{t|t}^e)\geq\lambda_{\min}(\mathrm{R})\|\Psi_t\|_{\mathrm{F}}^2 \geq\lambda_{\min}(\mathrm{R})\|\Psi_t\mathbf{v}\|^2. \tag*{(61)}

Step 4 (recursion for Ψt\Psi_t). For t>tct>t_c, the noises wt−1w_{t-1} and vtv_t are independent of η\eta, and the gains are deterministic for a given reception path. Multiplying (D.3) by η⊤\eta^\top, taking expectations and multiplying by H−1H^{-1} gives

Ψt=FtAΨt−1+KteMt−tc,Ψtc=PcC⊤H−1.\Psi_t = F_t^A\Psi_{t-1} + K_t^{\mathrm e}M^{t-t_c}, \qquad\Psi_{t_c} = P_cC^\top H^{-1}.

Let rt:=λ−(t−tc)Ψtvr_t := \lambda^{-(t-t_c)}\Psi_t v. Since Mt−tcv=λt−tcvM^{t-t_c}v = \lambda^{t-t_c}v,

rt=λ−1FtArt−1+Ktev.(62)r_t = \lambda^{-1}F_t^A r_{t-1} + K_t^{\mathrm e}v. \tag*{(62)}

Step 5 (limit). For t>tct > t_c, δte=1\delta_t^{\mathrm e}=1, so Πt∣t−1\Pi_{t|t-1} follows the standard Riccati recursion. By detectability and stabilizability, Πt∣t−1→Πˉ\Pi_{t|t-1} \to\bar{\Pi}, the stabilizing solution of the algebraic Riccati equation [43]. Hence Kte→Kˉ:=ΠˉC⊤(CΠˉC⊤+R)−1K_t^{\mathrm e} \to\bar{K} := \bar{\Pi}C^\top(C\bar{\Pi}C^\top+R)^{-1} and FtA→FˉAF_t^A \to\bar{F}_A, where Fˉ:=I−KˉC\bar{F}:=I-\bar{K}C and ρ(FˉA)<1\rho(\bar{F}_A)<1. Since ∣λ∣>1|\lambda|>1, λ−1FˉA\lambda^{-1}\bar{F}_A is Schur stable, and (D.5) is an asymptotically time-invariant, exponentially stable recursion with a convergent input. Therefore rt→rˉ:=(I−λ−1FˉA)−1Kˉvr_t \to\bar{r}:=(I-\lambda^{-1}\bar{F}_A)^{-1}\bar{K}v.

Step 6 (rˉ≠0\bar{r}\ne0). Since Πˉ⪰Q\bar{\Pi}\succeq Q and CQC⊤≻0CQC^\top\succ0, the matrix CKˉ=CΠˉC⊤(CΠˉC⊤+R)−1C\bar{K}=C\bar{\Pi}C^\top(C\bar{\Pi}C^\top+R)^{-1} is a product of two positive definite matrices and hence invertible. Thus Kˉv≠0\bar{K}v\ne0 and rˉ≠0\bar{r}\ne0.

Step 7 (conclusion). By Step 5, there is a finite T0T_0 such that ∥rt∥≥12∥rˉ∥\lVert r_t\rVert\ge\frac{1}{2}\lVert\bar{r}\rVert for t≥tc+T0t\ge t_c+T_0. Since ∥Ψtv∥=ρ(M)t−tc∥rt∥\lVert\Psi_t v\rVert=\rho(M)^{t-t_c}\lVert r_t\rVert, (D.4) gives

tr⁡(Pt∣te)≥cρ(M)2(t−tc),c:=14λmin⁡(R)∥rˉ∥2,\operatorname{tr}(P_{t|t}^{\mathrm e}) \ge c\rho(M)^{2(t-t_c)}, \qquad c:=\frac{1}{4}\lambda_{\min}(R)\lVert\bar{r}\rVert^2,

for all t≥tc+T0t\ge t_c+T_0. The constant cc depends only on (A,C,Q,R,M)(A,C,Q,R,M), not on tct_c. Since ρ(M)>1\rho(M)>1, tr⁡(Pt∣te)→∞\operatorname{tr}(P_{t|t}^{\mathrm e})\to\infty. ■\blacksquare

Proof of Theorem 3

The KLD between two Gaussian distributions N(μ1,Σ1)\mathcal{N}(\mu_1,\Sigma_1) and N(μ0,Σ0)\mathcal{N}(\mu_0,\Sigma_0) is given by [23]

Dg=12[tr⁡(Σ0−1Σ1)−p+(μ1−μ0)⊤Σ0−1(μ1−μ0)+log⁡det⁡Σ0det⁡Σ1].(63)D_g=\frac{1}{2}\left[\operatorname{tr}\left(\Sigma_0^{-1}\Sigma_1\right)-p+(\mu_1-\mu_0)^\top\Sigma_0^{-1}(\mu_1-\mu_0)+\log\frac{\det\Sigma_0}{\det\Sigma_1}\right]. \tag*{(63)}

We consider the case j=fj=f (ta≤τtt_a\le\tau_t). The densities comparable by the CUSUM test (18) are the conditional densities of the innovation given It−1\mathcal{I}_{t-1} and δtu=1\delta_t^{u}=1, see (19) and (22). Hence, on {δtu=1}\{\delta_t^{u}=1\}, μ0=0\mu_0=0, Σ0=Γt=CPt∣t−1C⊤+R\Sigma_0=\Gamma_t=CP_{t|t-1}C^\top+R from (22), μ1=μtf\mu_1=\mu_t^f from (20), and Σ1=Λktf\Sigma_1=\Lambda_{k_t}^f from (21). (The covariance δˉuΓt\bar{\delta}^{u}\Gamma_t in (A.2) is the unconditional covariance of δtuγt\delta_t^u\gamma_t, which is a Gaussian mixture with a point mass at 00 and is not the density used in the test.) These parameters depend on the past information It−1={γˉ1:t−1,e1:t−1,δ1:t−1}\mathcal{I}_{t-1}=\{\bar{\gamma}_{1:t-1},e_{1:t-1},\delta_{1:t-1}\}: μ1\mu_1 through the past innovations, the watermark and the last received falsified packet zˉτt\bar{z}_{\tau_t}, while Pt∣t−1P_{t|t-1} and ktk_t, and hence Λktf\Lambda_{k_t}^f, depend on the packet-reception history. Hence the conditional KLD D(fγ~tf∥fγt∣It−1)D(f_{\tilde{\gamma}_t^f}\mathbin\Vert f_{\gamma_t}\mid\mathcal{I}_{t-1}), given by DgD_g in (C.1), is a random variable, and, as in [23], we take its expectation over the joint distribution of It−1\mathcal{I}_{t-1}. Since μ1⊤Σ0−1μ1\mu_1^\top\Sigma_0^{-1}\mu_1 is a scalar, the cyclic property of the trace gives μ1⊤Σ0−1μ1=tr⁡(Σ0−1μ1μ1⊤)\mu_1^\top\Sigma_0^{-1}\mu_1=\operatorname{tr}(\Sigma_0^{-1}\mu_1\mu_1^\top). When δtu=0\delta_t^u=0 no packet is received, the statistic (17) is not updated, and the conditional KLD is zero by convention.

Since δtu\delta_t^u is independent of It−1\mathcal{I}_{t-1} and P{δtu=1}=δˉu\mathbb{P}\{\delta_t^u=1\}=\bar{\delta}^u, by linearity of the trace the trace and mean terms combine as

E[D(fγ~tf,fγt∣It−1)]=δˉu2E[tr⁡(Γt−1(Λktf+μ1μ1⊤)−p+log⁡det⁡Γtdet⁡Λktf)].(64)\mathbb{E}\left[D\left(f_{\tilde{\gamma}_t^f},f_{\gamma_t}\mid\mathcal{I}_{t-1}\right)\right] =\frac{\bar{\delta}^{u}}{2}\mathbb{E}\left[\operatorname{tr}\left(\Gamma_t^{-1}\left(\Lambda_{k_t}^f+\mu_1\mu_1^\top\right)-p+\log\frac{\det\Gamma_t}{\det\Lambda_{k_t}^f}\right)\right]. \tag*{(64)}

Mean term. By (20), we decompose μ1=ξt+Aaktzˉτt−CBet−1\mu_1=\xi_t+A_a^{k_t}\bar{z}_{\tau_t}-CBe_{t-1}, where ξt=Mktytd−C(A+BL)xˉt−1\xi_t=M^{k_t}y_t^d-C(A+BL)\bar{x}_{t-1}. The terms of ξt\xi_t depend on the watermark only up to et−2e_{t-2}, the attack process is independent of the watermark, and Γt\Gamma_t and ktk_t depend only on the packet-reception history. Since et−1e_{t-1} is zero mean and independent of ξt\xi_t, AaktzˉτtA_a^{k_t}\bar{z}_{\tau_t} and Γt\Gamma_t, the cross terms vanish, and, with N:=CBΣeB⊤C⊤⪰0N:=CB\Sigma_eB^\top C^\top\succeq0,

E[tr⁡(Γt−1μ1μ1⊤)]=E[tr⁡(Γt−1(ξt+Aaktzˉτt)(ξt+Aaktzˉτt)⊤)]+E[tr⁡(Γt−1N)]≥E[tr⁡(Γt−1N)].(65)\begin{aligned} \mathbb{E}\left[\operatorname{tr}\left(\Gamma_t^{-1}\mu_1\mu_1^\top\right)\right] &=\mathbb{E}\left[\operatorname{tr}\left(\Gamma_t^{-1}(\xi_t+A_a^{k_t}\bar{z}_{\tau_t})(\xi_t+A_a^{k_t}\bar{z}_{\tau_t})^\top\right)\right]\\ &\quad+\mathbb{E}\left[\operatorname{tr}\left(\Gamma_t^{-1}N\right)\right]\ge\mathbb{E}\left[\operatorname{tr}\left(\Gamma_t^{-1}N\right)\right]. \tag*{(65)} \end{aligned}

Hence, the watermark guarantees a non-zero mean shift under attack, even when the decoding reference zˉτt\bar{z}_{\tau_t} is falsified.

Bounds on Λktf\Lambda_{k_t}^f. Since δtu=1\delta_t^u=1, we have kt≥1k_t\ge1. From (21), Λktf=Qa+∑i=0kt−1AaiQa(Aai)⊤⪰Qa\Lambda_{k_t}^f=Q_a+\sum_{i=0}^{k_t-1}A_a^iQ_a(A_a^i)^\top\succeq Q_a, and, since ρ(Aa)<1\rho(A_a)<1, ∑i=0kt−1AaiQa(Aai)⊤⪯∑i=0∞AaiQa(Aai)⊤=Σz\sum_{i=0}^{k_t-1}A_a^iQ_a(A_a^i)^\top\preceq\sum_{i=0}^{\infty}A_a^iQ_a(A_a^i)^\top=\Sigma_z. Therefore, for every realization of ktk_t,

Qa⪯Λktf⪯Σz.(66)Q_a\preceq\Lambda_{k_t}^f\preceq\Sigma_z. \tag*{(66)}

Trace term. Substituting (C.3) into (C.2), and using the left inequality in (C.4) together with tr⁡(X−1Y)≥0\operatorname{tr}(X^{-1}Y)\ge0 for X≻0X\succ0, Y⪰0Y\succeq0, gives E[tr⁡(Γt−1(Λktf+N))]≥E[tr⁡(Γt−1(Qa+N))]\mathbb{E}[\operatorname{tr}(\Gamma_t^{-1}(\Lambda_{k_t}^f+N))]\ge\mathbb{E}[\operatorname{tr}(\Gamma_t^{-1}(Q_a+N))], in which the second argument is now deterministic. Since X↦tr⁡(X−1Y)X\mapsto\operatorname{tr}(X^{-1}Y) is convex and nonincreasing (in the Loewner order) on X≻0X\succ0 for Y⪰0Y\succeq0, Jensen’s inequality and, in steady state, E[Γt]=CE[Pt∣t−1]C⊤+R⪯CPˉC⊤+R\mathbb{E}[\Gamma_t]=C\mathbb{E}[P_{t|t-1}]C^\top+R\preceq C\bar{P}C^\top+R [41] yield

E[tr⁡(Γt−1(Λktf+N))]≥tr⁡(E[Γt]−1(Qa+N))≥tr⁡((CPˉC⊤+R)−1(Qa+N)).(67)\mathbb{E}\left[\operatorname{tr}\left(\Gamma_t^{-1}(\Lambda_{k_t}^f+N)\right)\right]\ge\operatorname{tr}\left(\mathbb{E}[\Gamma_t]^{-1}(Q_a+N)\right)\ge\operatorname{tr}\left((C\bar{P}C^\top+R)^{-1}(Q_a+N)\right). \tag*{(67)}

Log-determinant term. By the right inequality in (C.4), log⁡det⁡Λktf≤log⁡det⁡Σz\log\det\Lambda_{k_t}^f\le\log\det\Sigma_z for every realization of ktk_t. For log⁡det⁡Γt\log\det\Gamma_t, a lower bound is required. Since log⁡det⁡\log\det is concave, Jensen’s inequality only gives the upper bound E[log⁡det⁡Γt]≤log⁡det⁡E[Γt]\mathbb{E}[\log\det\Gamma_t]\le\log\det\mathbb{E}[\Gamma_t], so a bound on E[Pt∣t−1]\mathbb{E}[P_{t|t-1}] cannot be used here. Instead, we use an almost-sure lower bound on Pt∣t−1P_{t|t-1}. Let gδ(P):=APA⊤+Q−δAPC⊤(CPC⊤+R)−1CPA⊤g_\delta(P):=APA^\top+Q-\delta APC^\top(CPC^\top+R)^{-1}CPA^\top, so that, by (11)–(14), Pt∣t−1=gt−1δu(Pt−1∣t−2)P_{t|t-1}=g_{t-1}^{\delta^u}(P_{t-1|t-2}). For every P⪰0P\succeq0, g1(P)⪯g0(P)g_1(P)\preceq g_0(P), and g1g_1 is monotone in the Loewner order [41]. Let Pt∣t−1∗:=g1(Pt−1∣t−2∗)P_{t|t-1}^*:=g_1(P_{t-1|t-2}^*) with P1∣0∗P_{1|0}^* be the lossless Riccati recursion. By induction, Pt∣t−1=gt−1δu(Pt−1∣t−2)⪰g1(Pt−1∣t−2)⪰g1(Pt−1∣t−2∗)=Pt∣t−1∗P_{t|t-1}=g_{t-1}^{\delta^u}(P_{t-1|t-2})\succeq g_1(P_{t-1|t-2})\succeq g_1(P_{t-1|t-2}^*)=P_{t|t-1}^* for every packet-reception realization. Under the detectability and stabilizability assumptions, Pt∣t−1∗→P∗P_{t|t-1}^*\to P^*, the stabilizing solution of the standard ARE [43]. Hence, in steady state,

马会. Γt⪰CP⋆C⊤+R\Gamma_t \succeq C P^\star C^\top+ R almost surely, and by monotonicity of log det,

E[log⁡det⁡Γtdet⁡Λkt]≥log⁡∣CP⋆C⊤+R∣∣Σz∣.(68)\mathbb{E}\left[\log\frac{\det\Gamma_t}{\det\Lambda_{k_t}}\right] \ge\log\frac{\left|C P^\star C^\top+ R\right|}{\left|\Sigma_z\right|}. \tag*{(68)}

Applying (C.5) and (C.6) in (C.2) yields

E[D(fγ~tf,fγt∣It−1)]≥δˉu2(tr⁡((CPˉC⊤+R)−1(Qa+N))−p+log⁡∣CP⋆C⊤+R∣∣Σz∣),\mathbb{E}\left[D\left(f_{\tilde{\gamma}_t}^{f},f_{\gamma_t}\mid I_{t-1}\right)\right] \ge\frac{\bar{\delta}^{u}}{2}\left(\operatorname{tr}\left(\left(C\bar{P}C^\top+R\right)^{-1}(Q_a+N)\right)-p+\log\frac{\left|CP^\star C^\top+R\right|}{\left|\Sigma_z\right|}\right),

which is (28). ■

References

References

  1. [1]H. Sandberg, S. Amin, and K. H. Johansson, “Cyberphysical security in networked control systems: An introduction to the issue,” IEEE Control Systems Magazine, vol. 35, no. 1, pp. 20–23, 2015.DOI
  2. [2]A. A. Cárdenas, S. Amin, and S. Sastry, “Research challenges for the security of control systems,” HotSec, vol. 5, no. 15, p. 1158, 2008.
  3. [3]H. Sandberg, V. Gupta, and K. H. Johansson, “Secure networked control systems,” Annual Review of Control, Robotics, and Autonomous Systems, vol. 5, no. 1, pp. 445–464, 2022.DOI
  4. [4]B. Sinopoli, L. Schenato, M. Franceschetti, K. Poolla, and S. Sastry, “Lqg control with missing observation and control packets,” IFAC Proceedings Volumes, vol. 38, no. 1, pp. 1–6, 2005.
  5. [5]W. Liu, D. E. Quevedo, Y. Li, K. H. Johansson, and B. Vucetic, “Remote state estimation with smart sensors over markov fading channels,” IEEE Transactions on Automatic Control, vol. 67, no. 6, pp. 2743–2757, 2021.arxiv.org/abs/2005.07871
  6. [6]S. M. Dibaji, M. Pirani, D. B. Flamholz, A. M. Annaswamy, K. H. Johansson, and A. Chakrabortty, “A systems and control perspective of cps security,” Annual reviews in control, vol. 47, pp. 394–411, 2019.DOI
  7. [7]A. Humayed, J. Lin, F. Li, and B. Luo, “Cyber-physical systems security—a survey,” IEEE Internet of Things Journal, vol. 4, no. 6, pp. 1802–1831, 2017.arxiv.org/abs/1701.04525
  8. [8]M. S. Darup, A. B. Alexandru, D. E. Quevedo, and G. J. Pappas, “Encrypted control for networked systems: An illustrative introduction and current challenges,” IEEE Control Systems, vol. 41, no. 3, pp. 58–78, 2021.
  9. [9]J. Giraldo, A. A. Cardenas, and M. Kantarcioglu, “Security vs. privacy: How integrity attacks can be masked by the noise of differential privacy,” in 2017 American Control Conference (ACC). IEEE, 2017, pp. 1679–1684.
  10. [10]F. Farokhi and P. M. Esfahani, “Security versus privacy,” in 2018 IEEE Conference on Decision and Control (CDC). IEEE, 2018, pp. 7101–7106.DOI
  11. [11]X. Yan, G. Zhou, D. E. Quevedo, C. Murguia, B. Chen, and H. Huang, “Privacy-preserving state estimation in the presence of eavesdroppers: A survey,” IEEE Transactions on Automatic Science and Engineering, vol. 22, pp. 6190–6207, 2024.
  12. [12]J. Lee, K. Kapitanova, and S. H. Son, “The price of security in wireless sensor networks,” Computer Networks, vol. 54, no. 17, pp. 2967–2978, 2010.DOI
  13. [13]A. Tsiamis, K. Gatsis, and G. J. Pappas, “State-secrecy codes for networked linear systems,” IEEE Transactions on Automatic Control, vol. 65, no. 5, pp. 2001–2015, 2019.arxiv.org/abs/1709.04530
  14. [14]M. Lücke, J. Lu, and D. E. Quevedo, “Coding for secrecy in remote state estimation with an adversary,” IEEE Transactions on Automatic Control, vol. 67, no. 9, pp. 4955–4962, 2022.DOI
  15. [15]J. M. Kennedy, J. J. Ford, D. E. Quevedo, and F. Dressler, “Innovation-based remote state estimation secrecy with no acknowledgments,” IEEE Transactions on Automatic Control, vol. 69, no. 11, pp. 7433–7448, 2024.
  16. [16]M. J. Crimson, J. M. Kennedy, and D. E. Quevedo, “Remote state estimation with privacy against active eavesdroppers,” Automatica, vol. 171, p. 111932, 2025.arxiv.org/abs/2211.13411
  17. [17]P. Chen, W. Yang, Y. Liu, H. Zhang, and Z. Zhao, “Dynamic encoding scheme for state estimation over wireless sensor networks,” IEEE Transactions on Circuits and Systems II: Express Briefs, vol. 70, no. 11, pp. 4098–4102, 2023.
  18. [18]A. S. Leong, D. E. Quevedo, D. Dolz, and S. Dey, “Transmission scheduling for remote state estimation over packet dropping links in the presence of an eavesdropper,” IEEE Transactions on Automatic Control, vol. 64, no. 9, pp. 3732–3739, 2018.
  19. [19]A. Tsiamis, K. Gatsis, and G. J. Pappas, “State estimation with secrecy against eavesdroppers,” IFAC-PapersOnLine, vol. 50, no. 1, pp. 8385–8392, 2017.arxiv.org/abs/1612.04942
  20. [20]L. Wang, X. Cao, B. Sun, H. Zhang, and C. Sun, “Optimal schedule of secure transmissions for remote state estimation against eavesdropping,” IEEE Transactions on Industrial Informatics, vol. 17, no. 3, pp. 1987–1997, 2020.
  21. [21]A. S. Leong, A. Redder, D. E. Quevedo, and S. Dey, “On the use of artificial noise for secure state estimation in the presence of eavesdroppers,” 2018.
  22. [22]F. Pasqualetti, F. Dörfler, and F. Bullo, “Attack detection and identification in cyber-physical systems,” IEEE transactions on automatic control, vol. 58, no. 11, pp. 2715–2729, 2013.arxiv.org/abs/1202.6144
  23. [23]A. Naha, A. Teixeira, A. Ahlen, and S. Dey, “Quickest physical watermarking-based detection of measurement replacement attacks in networked control systems,” European Journal of Control, vol. 71, p. 100804, 2023.
  24. [24]W. Duo, S. Wang, and M. Zhou, “Multistep sensor attackability in cyber-physical systems,” IEEE Transactions on Automatic Control, vol. 70, no. 7, pp. 4794–4800, 2025.DOI
  25. [25]Y. Mo and B. Sinopoli, “Secure control against replay attacks,” in 2009 47th annual Allerton conference on communication, control, and computing (Allerton). IEEE, 2009, pp. 911–918.DOI
  26. [26]Y. Mo, S. Weerakkody, and B. Sinopoli, “Physical authentication of control systems: Designing watermarked control inputs to detect counterfeit sensor outputs,” IEEE Control Systems Magazine, vol. 35, no. 1, pp. 93–109, 2015.
  27. [27]B. Satchidanandan and P. R. Kumar, “Dynamic watermarking: Active defense of networked cyber–physical systems,” Proceedings of the IEEE, vol. 105, no. 2, pp. 219–240, 2016.arxiv.org/abs/1606.08741
  28. [28]H. Liu, Y. Mo, and K. H. Johansson, “Enhanced detection against replay attack: A survey on watermark design for cyber-physical systems,” in Safety, Security and Privacy for Cyber-Physical Systems. Springer, 2021, pp. 145–171.
  29. [29]D. Zhao, Y. Shi, S. X. Ding, Y. Li, and F. Fu, “Replay attack detection based on parity space method for cyber-physical systems,” IEEE Transactions on Automatic Control, vol. 70, no. 4, pp. 2390–2405, 2024.
  30. [30]X. Zhao, L. Liu, W. Xing, and N. Xu, “Stochastic event-based physical watermarks against replay attacks in cyber-physical systems,” IEEE Transactions on Control of Network Systems, vol. 11, no. 2, pp. 1035–1045, 2023.
  31. [31]H. Liu, Y. Zhang, Y. Li, and B. Niu, “Proactive attack detection scheme based on watermarking and moving target defense,” Automatica, vol. 155, p. 111163, 2023.
  32. [32]T. Li, Z. Wang, L. Zou, B. Chen, and L. Yu, “A dynamic encryption–decryption scheme for replay attack detection in cyber-physical systems,” Automatica, vol. 151, p. 110926, 2023.
  33. [33]H. Luo, S. Wu, and L. Li, “A time-domain estimation of the ν-gap metric and stability margin with its application to attack defense in cyber-physical systems,” Automatica, vol. 182, p. 112549, 2025.
  34. [34]G. Chen, X. Liu, X. Ren, J. Xia, and J. H. Park, “Asynchronous-based countermeasure for stealthy attack on aperiodic sampled-data control systems,” IEEE Transactions on Automatic Control, vol. 70, no. 9, pp. 6159–6166, 2025.
  35. [35]A. Tartakovsky, I. Nikiforov, and M. Basseville, Sequential analysis: Hypothesis testing and changepoint detection. CRC press, 2014.
  36. [36]A. Naha, A. M. Teixeira, A. Ahlen, and S. Dey, “Quickest detection of deception attacks on cyber–physical systems with a parsimonious watermarking policy,” Automatica, vol. 155, p. 111147, 2023.
  37. [37]A. Naha, A. Teixeira, A. Ahlen, and S. Dey, “Sequential detection of replay attacks,” IEEE Transactions on Automatic Control, vol. 68, no. 3, pp. 1941–1948, 2022.arxiv.org/abs/2012.10748
  38. [38]D. Sun, M. Cho, and I. Hwang, “On attack detection and identification for the cyber–physical system using lifted system model,” Automatica, vol. 176, p. 112243, 2025.
  39. [39]S. Weerakkody, O. Ozel, and B. Sinopoli, “A bernoulli-gaussian physical watermark for detecting integrity attacks in control systems,” in 2017 55th Annual Allerton Conference on Communication, Control, and Computing (Allerton). IEEE, 2017, pp. 966–973.arxiv.org/abs/1710.01105
  40. [40]J. Wu, Y. Li, D. E. Quevedo, V. Lau, and L. Shi, “Data-driven power control for state estimation: A Bayesian inference approach,” Automatica, vol. 54, pp. 332–339, 2015.arxiv.org/abs/1503.07158
  41. [41]L. Schenato, B. Sinopoli, M. Franceschetti, K. Poolla, and S. S. Sastry, “Foundations of control and estimation over lossy networks,” Proceedings of the IEEE, vol. 95, no. 1, pp. 163–187, 2007.DOI
  42. [42]Y.-C. Liu, G. Bianchin, and F. Pasqualetti, “Secure trajectory planning against undetectable spoofing attacks,” Automatica, vol. 112, p. 108655, 2020.arxiv.org/abs/1902.10869
  43. [43]B. D. O. Anderson and J. B. Moore, Optimal Filtering. Englewood Cliffs, NJ: Prentice-Hall, 1979.

Paper details

vlnc:W_4s7xj85w
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 10

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Engineering

Not recorded for this paper

View statement
Looking for related papers…
Contents