In classical game theory, security modeling has long relied on a comforting mathematical simplification: the zero-sum assumption. This framework dictates that an adversarial interaction is a closed economic loop where the defender's loss perfectly mirrors the attacker's gain. If a guard successfully protects a port, a checkpoint, or a server, the defender wins and the attacker loses an equivalent amount of utility.
This assumption is computationally elegant. It allows algorithms to leverage the minimax theorem, reducing complex strategic interactions to highly efficient linear programs. However, real-world security scenarios rarely operate as perfect mirrors. Over the past two decades, researchers have increasingly demonstrated that treating security as a zero-sum game creates a dangerous blind spot, leading to suboptimal resource allocation and vulnerable defense strategies.
The Illusion of the Perfect Mirror
Why does the zero-sum assumption fail? In practice, the motivations, costs, and valuations of defenders and attackers are fundamentally asymmetric. As Jason Tsai et al. observe in Urban Security: Game-Theoretic Resource Allocation in Networked Domains, assuming diametrically opposed payoffs fails to capture the complex, non-complementary incentives that drive real-world actors.
We can categorize these non-zero-sum discrepancies into three core frictions:
1. Asymmetric Valuations
The intrinsic value a defender places on an asset often has no relation to the utility an attacker derives from breaching it. Consider environmental protection. In Efficient Resource Allocation for Protecting Coral Reef Ecosystems, Bo An et al. point out that the ecological devastation of a destroyed marine ecosystem cannot be mapped directly to the black-market financial profit extracted by a poacher. The two parties are playing with entirely different currencies. Similarly, in network interdiction games, the cost to a state of allowing contraband through a border checkpoint does not scale linearly with the smuggler's financial payoff (Bo An et al., Optimal Interdiction of Illegal Network Flow).
2. Unreciprocal Operational Costs
Deploying security defenses is not free. It introduces friction, latency, and resource consumption. In Information Leakage Games, Mário S. Alvim et al. demonstrate that defensive measures—such as increased encryption or auditing overhead—penalize the defender in terms of power and latency without providing any corresponding benefit to the attacker. Conversely, an attacker faces independent computational expenses and implementation risks that do not directly translate into defender utility (Sarath Yasodharan et al., Nonzero-sum Adversarial Hypothesis Testing Games).
3. Mutual Losses and Rational Profit Maximization
In cyber-physical systems, adversarial actions can result in the destruction of shared infrastructure, yielding a catastrophic mutual loss rather than a win-lose outcome (Andrew Craig Cullen et al., Suboptimality of Constrained Action Adversarial Cyber-Physical Games). Furthermore, in decentralized economic networks like blockchains, attackers are often rational agents seeking to maximize their own financial returns through complex strategies like block withholding, rather than malicious actors whose sole purpose is to minimize the defender's utility (Krishnendu Chatterjee et al., Ergodic Mean-Payoff Games for the Analysis of Attacks in Crypto-Currencies).
The Cost of Misalignment
What happens when a defender optimizes for a zero-sum game that doesn't exist?
In a zero-sum model, the defender's strategy is entirely dictated by the attacker's potential gains. If the defender misjudges the attacker's true utility function, the calculated "optimal" defense targets a ghost incentive structure.
Nika Haghtalab et al. quantified this vulnerability in Lazy Defenders Are Almost Optimal against Diligent Attackers. They proved that while defenders can remain "lazy" and maintain near-optimal protection under strict zero-sum conditions even when facing limited surveillance, this robustness completely collapses in general-sum settings. When payoffs are asymmetric, ignoring the nuances of the attacker's actual motivations and information constraints causes a severe drop-off in the defender's real-world utility.
Zero-Sum Assumption:
[Defender Loss] <=======================> [Attacker Gain] (Perfect Mirror)
General-Sum Reality:
[Defender Utility] ---> Affected by: Internal Costs, Latency, Ecological Value
[Attacker Utility] ---> Affected by: Black-Market Profit, Execution Costs, Risk
This misalignment becomes even more pronounced when targets are interdependent. In critical infrastructure networks, an attack on a single node can trigger a cascade of failures across the entire system. Fang Liu demonstrates in Non-Additive Security Games that when the defender and attacker have fundamentally different valuation functions for these interconnected targets, the strategically zero-sum property breaks entirely. Linear solvers cannot find a stable equilibrium because the total utility is no longer a simple sum of independent target values.
Algorithmic Frontiers in General-Sum Modeling
If general-sum models are necessary, why did the field resist them for so long? The answer is computational complexity. As Bo An establishes in Infinite Horizon Adversarial Patrolling on Networks, transitioning from zero-sum to general-sum formulations strips away the efficiency of linear programming, introducing massive computational overhead that typically requires Non-Linear Programming (NLP) or Mixed-Integer Linear Programming (MILP).
To bridge this gap and make non-zero-sum modeling operationalizable, recent research focuses on scaling algorithms through two primary mechanisms:
Iterative Strategy Generation (Double Oracle)
Instead of computing the entire, intractable game matrix upfront, modern solvers use game-theoretic frameworks to dynamically expand the action space. For instance, in green security domains, Yi Wu et al. (Deep Reinforcement Learning for Green Security Games with Real-Time Information) combined the Double Oracle framework with Deep Reinforcement Learning (DRL). The system iteratively proposes best-response patrol routes based on real-time environmental feedback, finding general-sum equilibria without exploring the full exponential strategy space.
Constrained Reinforcement Learning
To handle complex operational trade-offs—such as minimizing passenger wait times while maintaining strict security thresholds—researchers have turned to multi-agent reinforcement learning with hard constraints. Arunesh Sinha et al. demonstrated in Solving Online Threat Screening Games using Constrained Action Space Reinforcement Learning that framing threat screening as a non-zero-sum game allows deep learning agents to optimize multi-faceted defender utilities that traditional minimax models simply cannot represent.
By abandoning the zero-sum illusion, these frameworks allow security agencies to move past rigid, reactive postures and design strategies that reflect the true economic, ecological, and operational realities of adversarial conflict.
Key Foundational Work
Urban Security: Game-Theoretic Resource Allocation in Networked Domains (2010) — contributed 1 concept elements to this direction.
Stochastic Games for Security in Networks with Interdependent Nodes (2010) — contributed 1 concept elements to this direction.
Infinite Horizon Adversarial Patrolling on Networks. (2011) — contributed 1 concept elements to this direction.