Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Extreme vulnerability to intruder attacks destabilizes network dynamics

Amirhossein Nazerian, Sahand Tangerami, Malbor Asllani, David Phillips, Hernan Makse, Francesco Sorrentino

TL;DR

The paper reveals a fundamental vulnerability of complex networks: a single intruder node, via adversarial connections, can destabilize network dynamics such as consensus, synchronization, and formation control. It introduces a transient-stability framework grounded in the algebraic connectivity $f(L)$ and transverse reactivity, proving that the worst-case attack concentrates the entire budget on one low-indegree node, a counterintuitive finding relative to hub-focused intuition. The results, proven for linear consensus and extended to nonlinear dynamics including Kuramoto models, yield universal scaling laws showing larger networks are, on average, more robust to single-node attacks under unidirectional coupling. These insights have broad practical implications for cyber-physical systems, power grids, sensor networks, and even ecological and social networks, and point toward defense strategies like attacker disconnection and targeted load shedding.

Abstract

Consensus, synchronization, formation control, and power grid balance are all examples of virtuous dynamical states that may arise in networks. Here we focus on how such states can be destabilized from a fundamental perspective; namely, we address the question of how one or a few intruder agents within an otherwise functioning network may compromise its dynamics. We show that a single adversarial node coupled via adversarial couplings to one or more other nodes is sufficient to destabilize the entire network, which we prove to be more efficient than targeting multiple nodes. Then, we show that concentrating the attack on a single low-indegree node induces the greatest instability, challenging the common assumption that hubs are the most critical nodes. This leads to a new characterization of the vulnerability of a node, which contrasts with previous work, and identifies low-indegree nodes (as opposed to the hubs) as the most vulnerable components of a network. Our results are derived for linear systems but hold true for nonlinear networks, including those described by the Kuramoto model. Finally, we derive scaling laws showing that larger networks are less susceptible, on average, to single-node attacks. Overall, these findings highlight an intrinsic vulnerability of technological systems such as autonomous networks, sensor networks, power grids, and the internet of things, with implications also to the realm of complex social and biological networks.

Extreme vulnerability to intruder attacks destabilizes network dynamics

TL;DR

The paper reveals a fundamental vulnerability of complex networks: a single intruder node, via adversarial connections, can destabilize network dynamics such as consensus, synchronization, and formation control. It introduces a transient-stability framework grounded in the algebraic connectivity and transverse reactivity, proving that the worst-case attack concentrates the entire budget on one low-indegree node, a counterintuitive finding relative to hub-focused intuition. The results, proven for linear consensus and extended to nonlinear dynamics including Kuramoto models, yield universal scaling laws showing larger networks are, on average, more robust to single-node attacks under unidirectional coupling. These insights have broad practical implications for cyber-physical systems, power grids, sensor networks, and even ecological and social networks, and point toward defense strategies like attacker disconnection and targeted load shedding.

Abstract

Consensus, synchronization, formation control, and power grid balance are all examples of virtuous dynamical states that may arise in networks. Here we focus on how such states can be destabilized from a fundamental perspective; namely, we address the question of how one or a few intruder agents within an otherwise functioning network may compromise its dynamics. We show that a single adversarial node coupled via adversarial couplings to one or more other nodes is sufficient to destabilize the entire network, which we prove to be more efficient than targeting multiple nodes. Then, we show that concentrating the attack on a single low-indegree node induces the greatest instability, challenging the common assumption that hubs are the most critical nodes. This leads to a new characterization of the vulnerability of a node, which contrasts with previous work, and identifies low-indegree nodes (as opposed to the hubs) as the most vulnerable components of a network. Our results are derived for linear systems but hold true for nonlinear networks, including those described by the Kuramoto model. Finally, we derive scaling laws showing that larger networks are less susceptible, on average, to single-node attacks. Overall, these findings highlight an intrinsic vulnerability of technological systems such as autonomous networks, sensor networks, power grids, and the internet of things, with implications also to the realm of complex social and biological networks.

Paper Structure

This paper contains 15 sections, 7 theorems, 50 equations, 6 figures.

Table of Contents

  1. Introduction
  2. RESULTS
  3. Balanced Directed Graphs
  4. General Directed Graphs
  5. Applications
  6. Conclusions
  7. Methods
  8. Transverse reactivity of stable and unstable consensus dynamics
  9. Proof of Proposition \ref{['prop1']}
  10. Proof of Proposition \ref{['prop2']}
  11. Proof of Proposition \ref{['prop3']}
  12. Proof of Proposition \ref{['prop4']}
  13. Digraphs with an Unidirectional Connection
  14. Digraphs with a Bidirectional Connection
  15. Proof of Mean-Slope Relation

Key Result

Proposition 1

Given $-c \leq 0$ and the Laplacian matrix $L_{aug}^b$ in Eq. eq:Lnewbi, the minimum algebraic connectivity $f$ is achieved when $b_{i^*} = -c$ for one node $i^*$ and $b_i = 0$ for all other nodes $i \neq i^*$, i.e., the entire budget must be allocated to one node.

Figures (6)

  • Figure 1: Illustration of an intruder attack on a network of drones attaining a given formation. The top panels show a network of coupled drones, before (a) and after (b) an intruder attack. The bottom panels display the trajectories of the drones before the attack in (c), showing convergence to the target positions, and after the attack in (d), showing divergence from the target position. The target positions are represented as stars in (c) and (d).
  • Figure 2: Schematic showing $f (L_{aug}^b)$ and $f (L_{aug}^u)$, the algebraic connectivity of asymmetric Laplacian matrices $L_{aug}^b$ and $L_{aug}^u$ in Eqs. \ref{['eq:Lnewbi']} (panel A: directed graphs with bidirectional adversarial connections) and Eq. \ref{['eq:Lnewuni']} (panel B: directed graphs with unidirectional adversarial connections), when only one node (either node $i$, $j$, or $k$) is pinned with the budget $-c < 0$. The dashed black line shows the line $f (L_{aug}^b) = -1.1c$ for panel A and shows the line $f (L_{aug}^u) = -0.1c$ for panel B. The colored dashed lines in panel A show $f (L_{aug}^b) = -2c + L_{xx}/2$ for $x = i, j,k$. The colored dashed lines in panel B show $f (L_{aug}^u) = -\frac{\sqrt{2}+1}{2} c + f_{inter}^x$ for $x = i, j,k$ where $f_{inter}^x = (3+2\sqrt{2})L_{xx}$. The schematics are not drawn in scale to enhance visualization.
  • Figure 3: The mean slope $< df / d c >$ for unidirectional (Eq. \ref{['eq:Lnewuni']}) connections of the adversarial agent for selected real networks with size $N$. The dashed black curve is the analytically predicted relation $< df / d c > = -1/N$.
  • Figure 4: Effect of adversarial agent addition in different applications. The top row panels (A-C) demonstrate the application of directly coupled chaotic oscillators, the middle row panels (D-F) depict the application of the power grid (nonlinear swing equation), and the bottom row panels (G-I) show the application of linear single-integrator robot formation control. The left column panels (A, D, G) show the topologies of their respective application. Panel A shows a directed network, and panels D and G show undirected networks. In all shown networks, black and orange links have a weight of $-1$ and $1$, respectively. The middle column panels (B, E, H) show the dynamics when the adversarial agent (denoted as node A) is not present, and the right column panels (C, F, I) show the dynamics when the adversarial agent is present. In panel D, the square blue nodes show the generators and the circle orange nodes are the load nodes (also the adversarial agent is a load node.) Panels B and C show the state $x_i (t)$ of the Lorenz oscillators; panels E and F show the angular position $\vartheta_i (t)$ of the loads/generators from the swing equation, and panels H and I show the position of the robots in $xy$ plane (diamonds label initial positions and stars label desired final positions.
  • Figure 5: Panel A shows a randomly generated scale-free network under attack through either of the colored nodes (blue $i=1$, orange $i=5$, yellow $i=17$, purple $i=94$, and green $i=100$). In panel B, we consider the dynamics of a network of coupled Lorenz oscillators, Eq. \ref{['eq:synch']}, and plot the norm of the transverse perturbations $\| \delta \pmb{X}(t) \|$ as a function of time $t$ and the attacked node $i$. Different curves correspond to different attacked nodes, all with the same budget $c = 1$. The dashed black line denotes the start time of the attack at $t = 4s$. The initial conditions of the oscillators are chosen to be the same in all cases where different nodes are attacked. The plot also provides the degree of node $i$ as $d_i$.
  • ...and 1 more figures

Theorems & Definitions (18)

  • Remark 1
  • Proposition 1
  • proof
  • Proposition 2
  • proof
  • Proposition 3
  • proof
  • Proposition 4
  • proof
  • Definition 1
  • ...and 8 more