EIPSIM: Modeling Secure IP Address Allocation at Cloud Scale

Eric Pauley, Kyle Domico, Blaine Hoak, Ryan Sheatsley, Quinn Burke, Yohan Beugin, Patrick D. McDaniel

Cloud computing relies on the efficient reuse of finite resources, most notably IPv4 addresses. When a virtual machine is decommissioned, its public IP is returned to a pool and often reassigned to a new tenant within minutes. This rapid churn creates a significant security vulnerability known as "latent configuration."

The Threat of Latent Configuration

Latent configuration occurs when external systems—such as DNS records, firewall allow-lists, or API configurations—continue to point to an IP address after its original owner has released it. An adversary can exploit this by "scanning" the cloud provider's IP pool, rapidly acquiring and releasing addresses to find one that still receives sensitive traffic or provides access to protected services.

Threat model showing how an adversary exploits IP recycling to capture traffic from previous tenants.

As illustrated in the threat model, an adversary can position themselves between two benign tenants. By cycling through IPs, the attacker captures "retrospective" traffic intended for Tenant 1 or "prospective" traffic intended for Tenant 2. Traditional defenses like "IP Tagging" (attempting to return the same IP to the same user) are vulnerable to Sybil attacks, where an attacker creates thousands of accounts to bypass per-account limits and "drain" the pool of fresh IPs.

IP Scan Segmentation

To address this, the researchers developed EIPSIM (Elastic IP Simulator) to test a novel defense: IP Scan Segmentation. This policy shifts the focus from who is requesting an IP to how they are using it. The core of the defense is behavioral profiling based on the mean allocation time ($\bar{d}_a$). Legitimate tenants typically hold IP addresses for long durations (hours or days), whereas attackers must cycle through them rapidly (minutes) to be effective.

The three-step process of the IP Scan Segmentation policy: Tracking tenant behavior, Logging IP history, and Decay of reputation over time.

The policy operates in three main stages:

  1. Tracking: The system maintains a running average of each tenant's allocation duration ($\bar{d}_a$).
  2. Logging: The system tracks the "cooldown" status ($t_{cd}$) of individual IPs based on their previous usage.
  3. Decay: A decay function ensures that the "taint" of a short-lived allocation eventually fades, allowing IPs to return to the general pool.

By using a cooldown heuristic, the algorithm prefers to assign IPs that were recently held for short periods to tenants who have a history of short-lived allocations. This effectively "quarantines" attackers into a small segment of the IP pool where they mostly recycle addresses among themselves, leaving the "clean" pool for legitimate users.

Evaluation and Results

The researchers validated EIPSIM using real-world server traces from Google (clusterdata-2019) and three years of "Internet Telescope" data. The results demonstrate that IP Scan Segmentation is highly effective:

  • Yield Reduction: It reduced the adversary's success rate in finding latent configurations by 83.8% compared to standard random allocation.
  • Superiority: It achieved a 70.1% improvement over the previous state-of-the-art "IP Tagging" defense.
  • Scalability: The simulation proved the algorithm could handle the massive throughput requirements of major cloud regions like AWS us-east-1, managing millions of IPs and thousands of requests per second.

Conclusion

IP Scan Segmentation provides a robust, Sybil-resistant defense against IP scanning in the cloud. By leveraging the fundamental behavioral difference between legitimate users and attackers—the duration of IP ownership—cloud providers can significantly improve the security of their infrastructure. While the "new tenant penalty" remains a challenge for those without a history, the policy represents a major step forward in securing the shared responsibility model of the cloud.

Summary updated

BibTeX
@inproceedings{pauley2025eipsim,
  author = {Eric Pauley and Kyle Domico and Blaine Hoak and Ryan Sheatsley and Quinn Burke and Yohan Beugin and Patrick D. McDaniel},
  title = {{EIPSIM: Modeling Secure IP Address Allocation at Cloud Scale}},
  booktitle = {NDSS},
  year = {2025},
}
Figure