Traditional game-theoretic security frameworks operate on a straightforward premise: the defender deploys a security policy, an attacker observes this policy, and a single-round interaction determines the outcome. This paradigm, foundational to classic Stackelberg Security Games (SSGs), has successfully optimized physical security resource allocation for over a decade, as surveyed by Sinha et al. (2018).
However, digital and cyber-physical domains break the core assumptions of static modeling. In cybersecurity, threats are rarely discrete, one-shot incidents. Instead, adversaries are persistent, adaptive agents who execute sequential moves, observe historical outcomes, and continuously alter their tactics to bypass defenses. To protect modern infrastructure, the research community is shifting from static, single-shot models to dynamic, multi-round, and sequential security games.
The Flaw of the One-Shot Assumption
When a game-theoretic model assumes a single-round equilibrium, it treats the environment as stationary. In reality, a smart attacker adapts to a defender's strategy over time. For instance, in machine learning security, standard classifiers assume data distributions remain constant, but Dritsoula et al. (2016) demonstrate that strategic attackers actively reshape their behavior to evade detection algorithms.
This mismatch between static models and adaptive adversaries spans multiple domains:
- Content and Email Filtering: Improving a spam or spear-phishing filter seems beneficial, but Laszka et al. (2015) note that attackers iteratively update their strategies to target an organization's shifting "weakest links." Furthermore, Ball et al. (2024) show that in multi-round filtering arms races, marginal technological improvements can be crowded out if human consumers decrease their vigilance or if attackers scale up their efforts.
- Network and Infrastructure Defense: In interdependent networks, an attacker does not just strike once; they execute a sequence of moves to navigate through a topology over time (Chan et al. (2015)). Similarly, in graph-based cleaning or pursuit scenarios, assuming static choices fails to capture environments where players move and react over multiple time steps (Akrida et al. (2019)).
- Insider Threats and Auditing: Static audit policies assume employees do not learn from previous interactions. Blocki et al. (2015) argue that modeling repeated games is essential because insiders adapt their behavior based on historical audit outcomes and observed patterns.
The Mechanics of Multi-Round Interaction
To capture the temporal evolution of threats, researchers model security as a sequential process where actions in the current round alter the state or information available in the next.
Instead of a single matrix game, interactions are framed as extensive-form games or repeated games. In wireless networks prone to backoff attacks—where malicious stations deviate from protocols to seize bandwidth—Parras et al. (2019) show that a repeated game framework allows players to achieve superior, stable cooperative equilibria that are impossible to sustain in a single-shot interaction.
Static Game: [Defender Policy] ──> [Attacker Response] ──> End
Dynamic Game: [Defender Move 1] ──> [Attacker Move 1] ──┐
│ (State & Knowledge Update)
[Defender Move 2] <── [Attacker Move 2] <─┘
This multi-stage reality changes how defenders must allocate resources across time. In cyber-physical systems like smart cities, static resource allocation cannot counter adversaries who adapt based on observed defender actions (Ferdowsi et al. (2017)). Real-time incidents, such as price tampering attacks in smart grids (Chen et al. (2021)) or malicious disruptions during power grid self-healing islanding procedures (Niu et al. (2021)), require dynamic frameworks where both parties adjust strategies as the incident unfolds.
Furthermore, threats may persist over an infinite horizon. Bonato et al. (2021) formalize this through the "Cops and Eternal Robbers" graph-pursuit game, demonstrating that defenders must be mathematically prepared to handle an endless sequence of sequential threats rather than terminating after a single capture. This continuous adaptation is also observed in economic security contexts, such as designing long-term government cybersecurity subsidies (An et al. (2019)), protecting drone delivery networks from persistent interdiction (Sanjab et al. (2017)), mitigating multi-round Sybil attacks in crypto token airdrops (Luo et al. (2025)), and countering multi-round misinformation campaigns in social networks (Bayiz et al. (2025)).
Information Asymmetry and Strategic Learning
A core feature of dynamic security modeling is managing how players learn. In a multi-round game, information itself is a utility and a strategic weapon. For example, in quantitative information flow scenarios, Alvim et al. (2017) and Alvim et al. (2020) model security as an information leakage game with non-linear utilities, demonstrating that players must dynamically adjust their strategies across repeated observations to optimize privacy protection.
Because attackers learn from historical data, defenders can use this learning process to their advantage through strategic deception and unpredictability:
- Evolving Deception: Schlenker et al. (2018) incorporate temporal dynamics into cyber deception, modeling how a defender can strategically alter network configurations over time to manipulate an attacker's evolving reconnaissance knowledge.
- Strategic Secrecy: While classic SSGs advocate for public commitment, An et al. (2017) introduce Disguised-resource Security Games (DSGs), proving that hiding resources (strategic secrecy) yields higher utility in multi-period encounters by preventing the attacker from easily mapping out defender patterns.
- Unpredictable Defenses: In federated learning, informed attackers can tailor malicious updates to bypass static aggregation rules. To counter this, the RobustTailor framework uses a mixed strategy Nash equilibrium across multiple rounds, keeping the server's defense unpredictable and neutralizing adaptive attacks (Pethick et al. (2024)).
- Payoff Manipulation: Rather than just moving guards, defenders can adaptively modify the attacker's payoff structure over time (e.g., changing fines or system values) to reshape the long-term dynamics of the competition (Shi et al. (2018)).
- Balanced Safeguarding: In deep neural network (DNN) backdoor security, Kallas et al. (2024) show that modeling interactions across multiple rounds of training and deployment reveals that the most effective long-term strategies rely on a calculated balance of power rather than maximum aggression.
Overcoming the Computational Explosion
If dynamic models are more realistic, why did the field rely on static models for so long? The answer lies in computational complexity. Moving from a single-shot game to a sequential game causes the state space to explode exponentially due to the branching nature of game trees.
To make dynamic modeling practical, researchers have developed specialized algorithmic transformations:
Sequence-Form Representations
Instead of converting an extensive-form game into a massive normal-form matrix, the authors of Bošanský et al. (2015) utilize a sequence-form algorithm. This approach expresses strategies as sequences of moves rather than combinations of pure actions, allowing for the efficient computation of Stackelberg equilibria in sequential games where players observe actions and update information over time.
Stochastic Game Formulations
When the environment changes state dynamically, researchers leverage discounted stochastic games. Singh et al. (2012) provided algorithms to compute Stackelberg equilibria in these settings, enabling defenders to maintain optimal policies even when attackers track and learn from defender patterns across shifting environmental states.
Approximation Schemes (FPTAS)
When exact equilibrium computation remains NP-hard, Fully Polynomial-Time Approximation Schemes (FPTAS) are used to find near-optimal strategies quickly. For instance, Blocki et al. (2015) introduced an optimization transformation that simplifies complex, restricted resource allocations in multi-resource audit games into a manageable mathematical form solved via an FPTAS. Similarly, Chan et al. (2015) provided an FPTAS to efficiently approximate Nash equilibria for attackers navigating tree-structured networks in interdependent defense scenarios.
As these algorithmic capabilities mature, benchmarks are evolving to match. Open-source frameworks like GUARD move beyond toy problems and random matrices, using real-world geospatial and demographic data to stress-test game-theoretic algorithms against highly complex, mobile, and dynamic adversaries (Krever et al. (2025)).
Key Foundational Work
Optimal Personalized Filtering Against Spear-Phishing Attacks (2015) — contributed 1 concept elements to this direction.
Audit Games with Multiple Defender Resources (2015) — contributed 1 concept elements to this direction.
Computing Stackelberg Equilibria in Discounted Stochastic Games (2012) — contributed 1 concept elements to this direction.