Introduction
Modern smart grids rely on state estimators to monitor system health and calculate real-time locational marginal prices (LMPs) for electricity markets. However, these systems are vulnerable to "data injection attacks," where malicious actors tamper with sensor measurements to create perceived grid congestion. This manipulation allows attackers to reap illicit financial profits through virtual bidding. While previous research has largely focused on single-attacker scenarios, this paper explores the more complex reality of multiple independent adversaries targeting the grid simultaneously.
Methodology: Game-Theoretic Modeling
The interaction between a grid operator (defender) and multiple attackers is modeled using two distinct game-theoretic frameworks:
- Stackelberg Game (Perfect Information): The defender acts as a leader, anticipating attacker responses to choose an optimal set of measurements to secure. The attackers, as followers, engage in a noncooperative game to reach a Generalized Nash Equilibrium (GNE).
- Hybrid Satisfaction-Nash Game (Limited Information): Recognizing that defenders often lack perfect information, this model seeks a "satisfaction equilibrium." Here, the defender aims for a strategy that keeps price deviations below a tolerable threshold rather than seeking absolute optimization.
To solve these games, the authors employ a distributed learning algorithm based on learning automata. This approach allows attackers to find optimal strategies based solely on their own payoff observations, without requiring knowledge of the grid topology or other adversaries.
Experimental Setup
The models were evaluated using the IEEE 30-bus test system, a standard benchmark for power system analysis. The simulation involved three independent attackers and one defender. Each attacker was assigned specific virtual bidding (VB) buses and target lines to manipulate, as detailed in the configuration below.
Key Insights and Results
The study reveals several critical dynamics in multi-adversary environments:
- Destructive Interference: Because the grid is a networked system, multiple attacks can cancel each other out. If two attackers attempt to create congestion in opposing directions, the net effect on the state estimate may be negligible.
- Strategic Defense ROI: The impact of defense is non-linear and highly efficient. While three unmitigated attackers caused a $306 financial impact on LMPs, defending just one strategically chosen measurement reduced this impact to $11.60. Defending two measurements using the Stackelberg strategy completely neutralized the attack ($0 impact).
- Stackelberg vs. Naive Defense: The game-theoretic approach significantly outperformed "critical defense" (protecting the most vulnerable lines). The naive approach allowed a 3% root mean square deviation (RMSD) in prices, whereas the Stackelberg defense prevented manipulation entirely.
- Price of Information (PI): The authors quantified the cost of limited information. When the defender used the Hybrid Satisfaction model to keep deviations under 5%, the PI was $5.74. This rose to $11.61 for a 10% threshold, illustrating the value of accurate attacker modeling.
Conclusion
This research demonstrates that while multiple adversaries complicate grid security, their conflicting interests can be leveraged by defenders. By using game-theoretic models to anticipate these interactions, grid operators can achieve near-total protection by securing a remarkably small number of sensor measurements. Future work will look to extend these models to non-linear AC power flow realities and explore the potential for attacker collusion.
Summary updated