Attack-Resilient Distributed Convex Optimization of Linear Multi-Agent Systems Against Malicious Cyber-Attacks over Random Digraphs

Continuous-Time Distributed Convex Optimization: paralleled with the discrete-time optimization schemes, the continuous-time distributed optimization algorithms have attracted much attention due to the well-developed continuous-time analysis techniques in control and the real-world cyber-physical system. In particular, the Zero-gradient-sum [5] and Newton-Raphson [6] algorithms are designed based on the positive and bounded Hessian of the local cost functions, while the Lagrangian-based algorithm based on the local gradients is adopted in [7]. The penalty-based optimization strategies are developed in [5] and [8], while the works investigate the multi-agent system with single-integrator dynamics only. The authors in [9] further consider distributed optimization of second-order multi-agent systems. Adaptive schemes are designed in [10] to achieve distributed optimization of linear agents via nonsmooth signum functions and gradients that satisfy special structures. All above designs require continuous communication. To remove this requirement, the time-based and event-triggered based strategies are presented in [11, 12, 13] over undirected and connected graphs. Besides, these algorithms need known initial states of each agent, which is difficult to verify in practice.

Resilient Distributed Convex Optimization: the closest related works on this topic have been recently published in [21, 22, 23, 24] in which the optimal solutions are obtained for first-order discrete-time multi-agent systems under Byzantine attacks (faulty nodes). In particular, the authors in [21] present the resilient optimization algorithm by removing $F$ (maximum amount of tolerable faults) nodes’ largest and smallest states at each iteration, such that the optimal solution converges to a convex hull of the set of all non-faulty nodes’ local minimizers. This algorithm is adopted in [22] to deal with constrained optimization problems. The local filtering consensus-based algorithms against $F$ faulty nodes are developed in [23], where the optimal solution is achieved under the $r$-robust condition. The modified algorithm named RDO-T is developed in [24], where the number of faulty agents is allowed to be any large. These aforementioned works assume that the faulty nodes and the removal of $F$ states are known a prior to the designer. Moreover, these algorithms rely on the network connectivity of the complete graph or the undirected and connected graph.

This paper is concerned with a resilient study of distributed optimization algorithms against cyber-attacks over random digraphs. The malicious DoS attacks, allowed to occur aperiodically, aim to interrupt information transmission. In addition to DoS attacks, the random digraphs are induced by unreliable networked constraints. The main contributions of this paper are summarized as follows. (1) inspired by our designs in [15, 17, 16, 18], this is the first to present time-based and event-based distributed optimization algorithms to achieve resilient distributed optimization of heterogeneous linear multi-agent networks against DoS attacks over random digraphs. The proposed algorithms that rely on a consensus-based gradient strategy with a switching system method to constrain DoS attacks, are capable of exactly seeking the optimal solutions under attacks over random digraphs. The global exponential convergence of the proposed algorithm can be ensured, provided that the frequency and duration of attacks satisfy certain bounded conditions; (2) these proposed resilient distributed optimization algorithms avoid continuous-time communication in many existing works (see [6, 7, 8, 5, 9, 10]). The proposed dynamic event-based distributed optimization scheme is proven to be free of Zeno behavior, and avoid fixed and periodic transmissions used in the time-based scheme; and (3) in contrast to related optimization works that require known positive Hessians of the local cost functions [5] and [6], known gradients satisfying special linear structures [10], known initial agent states [11, 12, 13], the proposed algorithms in this work remove those limitations to facilitate practical applications. Moreover, compared to optimization works in [21, 22, 23, 24] that consider Byzantine faults on nodes and require the removal of their states to be known a prior, these DoS attacks on communication networks are time-sequence based and allowed to occur aperiodically. Another contribution of this paper is that unlike works in [6, 7, 8, 5, 10, 11, 9, 12, 13] and [21, 22, 23, 24] over the undirected graph or weight-balanced digraph, the studied graphs are directed and time-varying, and under cyber-attacks, the graphs can be even disconnected or totally paralyzed.

Denote $\mathbb{R}$, $\mathbb{R}^{n}$, and $\mathbb{R}^{n\times m}$ as the sets of the real numbers, real $n$-dimensional vectors and real $n\times m$ matrices, respectively. Let $\mathbb{N}^{+}$ be the set of positive natural numbers. Let 0 (1) be the vector with all zeros (ones) with proper dimensions. Let col$(x_{1},...,x_{n})$ and diag$\{a_{1},...,a_{n}\}$ be a column vector with the entries $x_{i}$ and a diagonal matrix with the entries $a_{i}$, $i=1,2,\cdots,n$, respectively. $\otimes$ and $\left\|\cdot\right\|$ represent the Kronecker product and Euclidean norm, respectively. For a real matrix $M=M^{T}$, let $M>0$ be positive definite. Let $\lambda_{\min}(M)$, $\lambda_{\max}(M)$ be its minimum and maximum eigenvalues, respectively. Besides, $\sigma_{\max}(M)$ represents the maximum singular value of a matrix $M$. For a differentiable function $f:\mathbb{R}^{n}\rightarrow\mathbb{R}$, $\triangledown f$ is the gradient of $f$, and $f$ is strongly convex over a convex set $\mathbb{R}^{n}$, if $(x-y)^{T}(\triangledown f(x)-\triangledown f(y))>\iota||x-y||^{2}$ for $\forall x,y\in\mathbb{R}^{n},x\neq y$ and a scalar $\iota>0$. $f$ is locally Lipschitz at $x\in\mathbb{R}^{n}$ if there exists a neighbourhood $\mathcal{W}$ and a scalar $l$ so that $||f(x)-f(y)||\leq l||x-y||$ for $\forall x,y\in\mathcal{W}$; $f$ is locally Lipschitz on $\mathbb{R}^{n}$ if it is locally Lipschitz at $x$ for $\forall x\in\mathbb{R}^{n}$.

Static Digraph: let $\mathcal{G}$ $=$ $\left\{\mathcal{V},\mathcal{E}\right\}$ be a graph and $\mathcal{V}$ $\in$ $\left\{1,...,N\right\}$ be the set of vertices. The set of edges is denoted as $\mathcal{E}$ $\subseteq$ $\mathcal{V\times V}$. $\mathcal{N}_{i}$ $=$ $\left\{j\in\mathcal{V\mid}(j,i)\in\mathcal{E}\right\}$ is the neighborhood set of vertex $i$. For a directed graph $\mathcal{G}$, $(i,j)\in\mathcal{E}$ means that the information of node $i$ is accessible to node $j$, but not conversely. A matrix $A$ $=$ $\left[a_{ij}\right]$ is the adjacency matrix of $\mathcal{G}$, where $a_{ij}>0$ if $(j,i)\in\mathcal{E}$, else $a_{ij}=0$. A matrix $\mathcal{L}=[l_{ij}]$ is called the Laplacian matrix of $\mathcal{G}$, where $l_{ii}=\sum^{N}_{j=1}a_{ij}$ and $l_{ij}=-a_{ij}$, $i\neq j$.

Markovian Random Digraph: let $\mathcal{G}(t)=\left\{\mathcal{V},\mathcal{E}_{r(t)}\right\}$ be a time-varying digraph with $\mathcal{E}_{r(t)}$ being a set of edges, and $r(t):[0,\infty)\\ \rightarrow\mathcal{S}=\{1,2,...,s\}$ is a piecewise constant function with $\mathcal{S}$ being an index set of possible digraphs. The piecewise-constant function $r(t)$ is a Markovian signal. $\mathcal{A}(t)$ $=[a_{ij}^{r(t)}]$ is the adjacency matrix, where $a_{ij}^{r(t)}>0$ if $(j,i)\in\mathcal{E}_{r(t)}$, else $a_{ij}^{r(t)}=0.$ The neighboring set is denoted by $\mathcal{N}_{i}(t)=\left\{j\in\mathcal{V},(j,i)\in\mathcal{E}_{r(t)}\right\}$. Denote $\mathcal{L}(t)=[l^{r(t)}_{ij}]$, where $l^{r(t)}_{ii}=\sum^{N}_{j=1}a^{r(t)}_{ij}$ and $l^{r(t)}_{ij}=-a^{r(t)}_{ij}$, $i\neq j$.

Consider a multi-agent system consisting of $N$ agents governed by the following heterogeneous linear dynamics:

where $x_{i}\in\mathbb{R}^{n_{i}}$ denotes the state of agent $i$, $u_{i}\in\mathbb{R}^{p_{i}}$ denotes the control input of agent $i$, $y_{i}\in\mathbb{R}^{q}$ is its output, and $A_{i}\in\mathbb{R}^{n_{i}\times n_{i}}$, $B_{i}\in\mathbb{R}^{n_{i}\times p_{i}}$, $C_{i}\in\mathbb{R}^{q\times n_{i}}$ are constant system matrices. Suppose that the matrix pair $(A_{i},B_{i})$ is controllable and

General Distributed Optimization Problem: design a distributed scheme $u_{i}(t)$ for (1) using local interaction and information over a communication network so that the output of all agents can reach the optimal state $\theta^{*}$ that minimize: $\mathrm{F}(\theta)=\sum_{i=1}^{N}f_{i}(\theta),\ \theta\in\mathbb{R}^{q}$, where $f_{i}(\theta):\mathbb{R}^{q}\rightarrow\mathbb{R}$ is the private cost function known to agent $i$ only, and $\theta$ is the global decision variable to be optimized. From [7], it is equivalent to solve:

where $y_{i}\in\mathbb{R}^{q}$ is a local estimate on the optimal solution $\theta^{*}$, and $y=\text{col}(y_{1},\cdots,y_{N})$ is its stack vector of all estimates.

To solve the problem, two standard assumptions are introduced.

In large-scale cyber-physical systems, the wireless communication may not be reliable due to certain physical uncertainties, e.g., failures, quantization errors, and packet losses in a digital communication [25]. We consider an unreliable network consisting of $N$ agents whose integrated wireless communication links are time-varying and failure-prone with certain probabilities. As considered in [25], a random Markov chain model can be adopted to capture this situation. This random process describes dynamic changes of topologies. Let $r(t)$ be a right-continuous homogeneous Markovian process on the probability space taking values in a finite state space $\mathcal{S}=\{1,2,...,s\}$ with an infinitesimal generator $\Upsilon=(\gamma_{pq})$, given by $\mathrm{P}_{pq}(t)=\mathrm{\Pr ob}\{r(t+h)=q|r(t)=p\}=\gamma_{pq}h+o(h)$, if $p\neq q$, else, $1+\gamma_{pp}h+o(h)$, where $\gamma_{pq}\geq 0$ is the transition rate from the state $p$ to the state $q$, while $\gamma_{pp}=-\sum^{s}_{q=1,p\neq q}\gamma_{pq},$ and $o(h)$ satisfies: $\lim_{h\rightarrow 0}o(t)/h=0$.

As studied in [18], the DoS attack refers to the interruption of the communication channels. Without loss of generality, suppose that there exists an attack sequence $\{a_{m}\}_{m\in\mathbb{N}}$ when a DoS attack is lunched at $a_{m}$, and let the duration of this attack be $\tau_{m}\geq 0$. Then, the $m$-th DoS attack strategy can be generated with $\mathcal{A}_{m}$ $=$ ${a_{m}}\cup[a_{m},a_{m}+\tau_{m})$ with $a_{m+1}>a_{m}+\tau_{m}$ for all $m\in\mathbb{N}.$ Thus, for given $t\geq t_{0}$, the sets of time instants where communication is denied (unsuccessful) are described by

which implies that on the interval $[t_{0},t]$, the sets of time instants where communication is allowed are: $\Xi_{s}(t_{0},t)=[t_{0},t]\setminus\Xi_{a}(t_{0},t)$. In words, $|\Xi_{a}(t_{0},t)|$ and $|\Xi_{s}(t_{0},t)|$ represent the total lengths of the attacker being active and sleeping over $[t_{0},t]$, respectively.

The main objective of the paper is to solve a resilient distributed optimization issue of heterogeneous linear multi-agent systems.

Based on the DoS attack model in the subsection II-E, without loss of generality, suppose that there exists an infinite sequence of uniformly bounded non-overlapping time intervals $[t_{k},t_{k+1}),k=0,1,2\cdots$. Each agent $i$ thus updates its controller $u_{i}(t)$ over time intervals $[t_{k},t_{k+1})$. Then, in the presence of the DoS attacks, we denote the attack-induced new time sequence as $0=t_{0}=a_{0}<t_{1}<a_{1}<a_{1}+\tau_{1}\leq t_{2}<\cdots<t_{k}<a_{k}<a_{k}+\tau_{k}\leq t_{k+1}<\cdots$.

Next, we propose the following distributed algorithm to analyze the effect of DoS attacks over random digraphs.

where $\alpha,\beta\in\mathbb{R}$ are positive constant gains, $\vartheta_{i}$ is the intermediate variable, and $e_{\varrho zi}$, $e_{yi}$ are consensus errors under DoS attacks

where $\varrho_{i},z_{i}\in\mathbb{R}^{q}$ denote the auxiliary states. In (8), $K_{i}\in\mathbb{R}^{p_{i}\times n_{i}}$ and $(U_{i},W_{i},X_{i})$ is the solution to the following equations:

Define $\tilde{\mathcal{L}}(t)=\mathcal{L}(t)\otimes I_{q}$. Substituting (8)-(10) into (1) gives the following closed-loop system under attacks and random digraphs.

where $\chi=\text{col}(x,\varrho,z)$ and $x,\varrho,z,\vartheta,\triangledown\tilde{f}(y)$ are stacked vectors of $x_{i}$, $\varrho_{i}$, $z_{i}$, $\vartheta_{i}$, $\triangledown f_{i}(y_{i})$, and $A,B,C,K,U,W,X$ are stacked block diagonal matrices of $A_{i}$, $B_{i}$, $C_{i}$, $K_{i}$, $U_{i}$, $W_{i}$, $X_{i}$, respectively.

Next, we present the exponential resilient distributed optimization result against DoS attacks and random digraphs.