Identity Concealment Games: How I Learned to Stop Revealing and Love the Coincidences

Mustafa O. Karabag, Melkior Ornik, Ufuk Topcu

Introduction: The Challenge of Stealth

In traditional adversarial settings, the primary goal of a hostile agent is often simple: achieve a specific objective, such as compromising a server or evading capture. However, in many real-world scenarios, success is hollow if the agent is immediately identified. A Denial-of-Service (DoS) attacker who is instantly blocked, or a spy whose erratic behavior flags them to surveillance, fails their broader mission. This paper introduces "identity concealment games," a novel framework designed to model scenarios where a hostile agent must achieve a malicious goal while masquerading as a benign, "average" user.

Conceptual overview of an identity concealment scenario where an attacker mimics real client behavior to evade server detection.

Problem Definition

Standard game theory typically focuses on winning or losing in a zero-sum environment. It often ignores the cost of detection. If a defender plays suboptimally, standard reinforcement learning suggests the attacker should "explore" to find weaknesses. However, in stealth-critical environments, active exploration is dangerous; taking unusual actions to test a defender's response can immediately reveal the attacker's hostile intent. The core problem addressed here is how an agent can learn to exploit a defender's vulnerabilities without ever "breaking character" or deviating significantly from expected benign behavior.

Methodology: Formalizing Concealment

The authors formalize this problem as a two-player zero-sum stochastic reachability game. To quantify stealth, they utilize the Kullback-Leibler (KL) divergence. By defining an "average player" policy (representing normal, non-malicious behavior), the game's objective function is modified to penalize the hostile player whenever their action distribution deviates from this baseline.

A state transition diagram representing the stochastic reachability game model used to formalize identity concealment.

Equilibrium and Offline Learning

The research proves that an equilibrium policy pair exists for these games, establishing a mathematically optimal balance between objective achievement and stealth. To solve the exploration dilemma, the authors propose an offline learning algorithm. Instead of direct interaction, the hostile agent learns by observing a dataset of historical interactions between average players and the defender. This "parasitic learning" approach allows the agent to synthesize an attack policy entirely offline. The algorithm incorporates a safety mechanism that heavily penalizes entering "unknown" states—those lacking sufficient historical data—ensuring the agent remains within the bounds of observed, "safe" behavior.

Experimental Evaluation

The framework was evaluated using two distinct simulated environments:

  1. Cyber Interaction Scenario: A hostile client attempts a DoS attack on a server. The attacker must mimic the request patterns of legitimate clients to avoid being flagged by the server's detection systems.
  2. Pursuit-Evasion Game: An evader attempts to reach a goal while being pursued in a grid-based environment. The evader must mimic the movement patterns of a benign, wandering agent, only deviating when necessary to secure an escape.

Key Results

The theoretical contributions are backed by rigorous sample complexity bounds, proving how much historical data is required to guarantee a near-optimal policy. In the cyber simulation, attackers using the equilibrium policy were significantly harder to detect (as shown via ROC curves) than those using aggressive, greedy strategies. In the pursuit-evasion tests, the offline learning algorithm successfully exploited the pursuer's suboptimal habits, achieving a higher win rate than the theoretical equilibrium without ever alerting the pursuer during the learning phase.

Insights and Future Directions

A critical insight of this work is the framing of identity concealment as a statistical hypothesis testing problem. By minimizing KL divergence, the attacker effectively makes it statistically difficult for the defender to distinguish malicious intent from random coincidence. The optimal strategy is not to act exactly like a normal user at all times, but to act normally when possible and only "break character" when it is strategically essential.

While the current framework assumes a stationary baseline for "average" behavior, future work could explore non-stationary environments where user patterns drift over time. Additionally, extending these concepts to partially-observable environments (POMDPs) would further enhance the realism and applicability of identity concealment games in complex, real-world security domains.

Summary updated

BibTeX
@article{karabag2024identity,
  author = {Mustafa O. Karabag and Melkior Ornik and Ufuk Topcu},
  title = {{Identity Concealment Games: How I Learned to Stop Revealing and Love the Coincidences}},
  journal = {Automatica},
  year = {2024},
  url = {https://openreview.net/forum?id=YsXl3xTiyc},
}
Figure