Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Targeted Algorithmic Purpose-Driven Cyber Attacks in Distributed Multi-Agent Optimization

Mahan FakouriFard, Mingxi Liu

TL;DR

The paper addresses the vulnerability of distributed multi-agent optimization (DMAO) to targeted, stealthy algorithmic attacks. It formulates a rigorous framework for purpose-driven attacks that can be implemented in primal or dual modes, derives sharp bounds on how much attacked optima can deviate from nominal solutions, and proves an exact equivalence between dual and primal attack strategies. The authors instantiate the theory via a distributed electric-vehicle (EV) charging valley-fill problem, showing that attackers can steer charging profiles and system metrics (e.g., voltages) toward self-serving goals while preserving convergence and feasibility. Simulations on a 500-EV fleet validate the theoretical results, demonstrating both primal and dual attacks can achieve attack goals with measurable shifts in charging trajectories and minimal detectability. The work highlights a new undercover threat in DMAO for cyber-physical systems and motivates future defenses and detection mechanisms to mitigate such for-purpose manipulations.

Abstract

Distributed multi-agent optimization (DMAO) enables the scalable control and coordination of a large population of edge resources in complex multi-agent environments. Despite its great scalability, DMAO is prone to cyber attacks as it relies on frequent peer-to-peer communications that are vulnerable to malicious data injection and alteration. Existing cybersecurity research mainly focuses on \emph{broad-spectrum} attacks that aim to jeopardize the overall environment but fail to sustainably achieve specific or targeted objectives. This paper develops a class of novel strategic purpose-driven algorithmic attacks that are launched by participating agents and interface with DMAO to achieve self-interested attacking purposes. Theoretical foundations, in both primal and dual senses, are established for these attack vectors with and without stealthy features. Simulations on electric vehicle charging control validate the efficacy of the proposed algorithmic attacks and show the impacts of such attacks on the power distribution network.

Targeted Algorithmic Purpose-Driven Cyber Attacks in Distributed Multi-Agent Optimization

TL;DR

The paper addresses the vulnerability of distributed multi-agent optimization (DMAO) to targeted, stealthy algorithmic attacks. It formulates a rigorous framework for purpose-driven attacks that can be implemented in primal or dual modes, derives sharp bounds on how much attacked optima can deviate from nominal solutions, and proves an exact equivalence between dual and primal attack strategies. The authors instantiate the theory via a distributed electric-vehicle (EV) charging valley-fill problem, showing that attackers can steer charging profiles and system metrics (e.g., voltages) toward self-serving goals while preserving convergence and feasibility. Simulations on a 500-EV fleet validate the theoretical results, demonstrating both primal and dual attacks can achieve attack goals with measurable shifts in charging trajectories and minimal detectability. The work highlights a new undercover threat in DMAO for cyber-physical systems and motivates future defenses and detection mechanisms to mitigate such for-purpose manipulations.

Abstract

Distributed multi-agent optimization (DMAO) enables the scalable control and coordination of a large population of edge resources in complex multi-agent environments. Despite its great scalability, DMAO is prone to cyber attacks as it relies on frequent peer-to-peer communications that are vulnerable to malicious data injection and alteration. Existing cybersecurity research mainly focuses on \emph{broad-spectrum} attacks that aim to jeopardize the overall environment but fail to sustainably achieve specific or targeted objectives. This paper develops a class of novel strategic purpose-driven algorithmic attacks that are launched by participating agents and interface with DMAO to achieve self-interested attacking purposes. Theoretical foundations, in both primal and dual senses, are established for these attack vectors with and without stealthy features. Simulations on electric vehicle charging control validate the efficacy of the proposed algorithmic attacks and show the impacts of such attacks on the power distribution network.

Paper Structure

This paper contains 18 sections, 34 equations, 5 figures.

Table of Contents

  1. Introduction
  2. Problem Formulation and Preliminaries
  3. Distribution network and EV charging model
  4. Valley filling -- a cooperative multi-agent optimization
  5. Distributed EV charging control scheme
  6. Purpose-Driven Algorithmic Cyber Attacks
  7. Targeted algorithmic primal attack vectors
  8. Time tuning attack
  9. Battery damage attack
  10. Dual targeted algorithmic cyber-attack
  11. Simulation Results
  12. Primal attack scenarios
  13. Attack-free scenario
  14. Battery damage attack
  15. Dual attack scenarios
  16. ...and 3 more sections

Figures (5)

  • Figure 1: Charging profiles of all EVs under (a) No attack, (b) Primal battery damage, (c) Primal smooth-charging, (d) Dual smooth-charging, (e) Primal time-tuning, and (f) Dual time-tuning attack.
  • Figure 2: Baseline load and total load under different attacks.
  • Figure 3: Nodal voltage magnitudes. Star-marked lines represent the baseline case, and solid lines represent the case with controlled EV charging loads.
  • Figure 4: Charging profiles of the 3-EV toy problem under dual attack
  • Figure 5: Values of calculated vectors $\bm{\bar{\mathcal{P}}}_1$ for dual attack scenarios