On Dynamic Recovery of Cloud Storage System Under Advanced Persistent Threats

Pengdeng Li, Xiaofan YangIEEE Access 2019 · OpenReview

Advanced Persistent Threats (APTs) represent a significant challenge for cloud storage systems (CSSs). Unlike typical cyberattacks, APTs are stealthy, continuous, and highly strategic, often remaining undetected for months to exfiltrate valuable data. This paper, originally published in IEEE Access, introduces a rigorous mathematical framework to address the dynamic recovery of cloud servers while an APT actor is actively attempting to steal data.

The Dynamic Cloud Storage Recovery (DCSR) Problem

When an APT is detected, the cloud defender faces a critical balancing act known as the Dynamic Cloud Storage Recovery (DCSR) problem. The defender must allocate limited security resources—such as manpower, budget, and computational time—to clean and restore compromised servers.

If recovery is too slow, the financial impact of data loss grows exponentially. Conversely, if the recovery is too aggressive, the operational costs of the defense itself can become prohibitive. The goal is to find an optimal recovery rate that minimizes the total loss in an environment where the attacker is also adjusting their tactics in real-time.

The cloud defender manages recovery decisions across storage servers serving various clients.

Modeling the Conflict: Differential Game Theory

To solve the DCSR problem, the authors employ Differential Game Theory. This mathematical approach models strategic interactions that evolve over time. The conflict is framed as a game between two parties:

  1. The Attacker: Seeks to maximize their net benefit, calculated as the value of stolen data minus the costs incurred during the attack.
  2. The Defender: Seeks to minimize their total loss, which includes the value of lost data and the costs of recovery efforts.
An APT attacker uses a C&C server to compromise storage servers and exfiltrate data.

The researchers developed an Expected State Evolution Model, treating each storage server as a system that transitions between "uncompromised" and "compromised" states based on time-varying attack and recovery rates. By applying the Nash equilibrium concept, they identified a state where neither party can improve their outcome by unilaterally changing their strategy.

Key Insights: Strategy and Cost

The study reveals several critical insights into optimal defense strategies:

  • Front-Loading Strategy: In a competitive scenario, both the attacker and defender typically start with high-intensity actions that taper off. The attacker aims to exfiltrate data before defenses are fully established, while the defender seeks to "stop the bleeding" as quickly as possible.
  • Cost Function Geometry: The "shape" of the cost functions—how expensive it is to increase attack or recovery intensity—dictates the optimal strategy. If costs are concave, the strategy is "bang-bang" (switching abruptly between minimum and maximum rates). If costs are strictly convex, the rates transition smoothly over time.
  • Robustness: A significant finding is that the Nash equilibrium strategy protects the defender even against "dumb" or irrational attackers. If the attacker deviates from the optimal strategy, the defender’s total loss generally decreases.

Results and Scalability

The researchers validated their model using synthetic data and numerical simulations on environments with up to 100 storage servers. The derived Competitive Strategy Profile (CSP) consistently outperformed 100 randomly generated strategy profiles, resulting in lower total losses for the defender.

Furthermore, the approach proved computationally efficient. Unlike traditional game-theoretic solvers that suffer from the "curse of dimensionality," the authors' necessity-system approach scales effectively, making it a viable candidate for real-world cloud architectures.

Conclusion

While the current model assumes perfectly rational actors and full information, it provides a foundational roadmap for reactive defense in cloud environments. Future work may incorporate Bayesian games to account for the "fog of war" in cybersecurity or utilize Prospect Theory to model human psychological biases during crisis management. By moving beyond simple proactive defense, this research enhances the ability of cloud providers to maintain resilience under persistent, strategic threats.

Reference: Li, P., & Yang, X. (2019). On Dynamic Recovery of Cloud Storage System Under Advanced Persistent Threats. IEEE Access, 7, 103556-103569.

Summary updated

Abstract

Advanced persistent threat (APT) for data theft poses a severe threat to cloud storage systems (CSSs). An APT actor may steal valuable data from the target CSS even in a strategic fashion. To protect a CSS from APT, the cloud defender has to dynamically allocate the limited security resources to recover the compromised storage servers, aiming at mitigating his total loss. This paper addresses this dynamic cloud storage recovery (DCSR) problem by employing differential game theory. First, by introducing an expected state evolution model capturing the CSS's expected state evolution process under a combination of attack strategy and recovery strategy, we measure the APT attacker's net benefit and the cloud defender's total loss. On this basis and in the worst-case situation where the cloud defender assumes that the APT attacker has full knowledge of his expected loss, we reduce the DCSR problem to a differential game-theoretic problem (the DCSR* problem) to characterize the strategic interactions between the two parties. Second, we derive a necessary condition for Nash equilibrium of the DCSR* problem and thereby introduce the concept of competitive strategy profile. Next, we study the structural properties of the competitive strategy profile, followed by some numerical examples. Then, we conduct extensive comparative experiments to exhibit that the competitive strategy profile is superior to a large number of randomly generated strategy profiles in the sense of Nash equilibrium solution concept. Finally, we briefly analyze the practicability (scalability and feasibility) of this paper. Our findings will be helpful to enhance the APT defense capabilities of the cloud defender.
BibTeX
@inproceedings{li2019dynamic,
  author = {Pengdeng Li and Xiaofan Yang},
  title = {{On Dynamic Recovery of Cloud Storage System Under Advanced Persistent Threats}},
  booktitle = {IEEE Access},
  year = {2019},
  url = {https://openreview.net/forum?id=WZcMsybKTYa},
}
Figure