Heat-seeking honeypots: design and experience

John P. John, Fang Yu, Yinglian Xie, Arvind Krishnamurthy, Martín AbadiWWW 2011 · DOI

TL;DR

This paper introduces "heat-seeking honeypots," a novel approach to understanding web attacker behavior by dynamically generating decoy web pages that mimic those targeted by attackers. By attracting and analyzing attacker traffic, they demonstrate a low-cost, scalable method to identify attack patterns and develop effective defense strategies.

What Problem Does This Paper Address?

Compromised web servers are a significant problem, often used for malicious activities like phishing, malware distribution, and open proxies. Attackers actively seek out these vulnerable servers, leveraging their high pageranks and free resources. Understanding how attackers find, compromise, and exploit these servers is crucial for developing effective defenses. Traditional honeypots often require significant manual effort to set up and maintain, making them difficult to scale and adapt to the ever-evolving threat landscape. This paper aims to bridge this gap by creating a system that can automatically adapt to current attacker interests and attract them with minimal human intervention.

Methods

The core innovation is the "heat-seeking honeypot" system, which comprises four key components:

  1. Attacker Query Identification: The system leverages search engine logs (specifically Bing) to identify queries that attackers are likely using to find vulnerable servers. Tools like SBotMiner are used to group similar queries and identify patterns.
  2. Honeypot Page Creation: Instead of installing actual vulnerable software (which is resource-intensive and requires expert knowledge), the system dynamically generates web pages that look like they belong to vulnerable applications. This is achieved by fetching pages from search results for attacker queries and then encapsulating their HTML content within a simple PHP script that logs all interactions. This approach allows for rapid generation and deployment without needing to understand specific exploits.
  3. Honeypot Advertising: To increase visibility, the honeypot pages are submitted to search engines, and their pagerank is boosted through surreptitious links on other public web pages. This aims to ensure they appear in search results for relevant attacker queries.
  4. Attack Traffic Detection and Analysis: The system logs all incoming requests. It then employs a whitelist-based approach to distinguish malicious traffic from legitimate crawler or user visits. This involves identifying known crawler behavior and then flagging any requests for non-whitelisted resources as suspicious.

Datasets

The paper does not explicitly mention training datasets. However, it uses:

  • Bing Search Engine Logs: To identify attacker queries.
  • Web Server Access Logs: From 100 randomly selected public web servers for evaluation of the whitelist-based detection method.
  • Honeypot Logs: Generated by their deployed honeypots over a three-month period, containing over 54,000 visits from 6,400 distinct IP addresses.

Key Insights and Ideas

  • Dynamic Page Generation is Scalable: The ability to generate honeypot pages automatically based on attacker search queries is a significant advantage, allowing the system to adapt to new threats without manual intervention.
  • Mimicking Vulnerable Pages is Effective: Attackers are easily fooled by pages that closely resemble those of known vulnerable applications, even if no actual vulnerable software is running. This significantly reduces the operational overhead of honeypots.
  • Whitelist-Based Detection is Robust: A whitelist approach, identifying legitimate access patterns (like crawler behavior) and flagging anything else, proves to be a more general and automated method for detecting malicious traffic compared to traditional blacklists.
  • Attackers Exploit Common Vulnerabilities: While some popular vulnerabilities like XSS and SQL injection were less prevalent in their honeypot logs, file disclosure (FILE) and XMLRPC-related attacks were highly common, indicating specific attacker priorities.
  • Aggressive Attackers Exist: A small percentage of IP addresses accounted for a disproportionately large number of visits, suggesting a subset of attackers are highly persistent and employ brute-force tactics.

Key Results

  • Over a three-month period, their deployed honeypots attracted over 54,000 visits from nearly 6,400 distinct IP addresses.
  • The honeypot system successfully identified malicious queries and generated relevant pages, attracting traffic comparable to real vulnerable applications but at a much lower setup cost.
  • The whitelist-based detection method achieved a false-negative rate of at most 1%.
  • When applied to 100 regular web servers, the whitelist approach identified malicious traffic in all logs, with attackers sometimes accounting for 90% of the total non-crawler traffic.
  • The most popular honeypot page, attracting over 10,000 visits, was for a Joomla content management system.
  • File disclosure (FILE) and XMLRPC-related attacks constituted the largest portions of observed traffic (43.57% and 18.97% respectively).

Limitations and Future Work

  • Discoverability Delay: Heat-seeking honeypots rely on search engines indexing their pages, which introduces a delay between page creation and attacker discovery. This can be mitigated by linking to popular pages or through search engine cooperation.
  • Limited Interaction Depth: While effective at attracting attackers, the dynamically generated pages offer less interaction depth compared to high-interaction honeypots running actual software. This means attackers might not fully exploit vulnerabilities as they would on a real system.
  • Attacker Sophistication: The paper acknowledges that attackers might eventually detect the honeypot if they perform deep analysis.
  • Future Work: The authors suggest extending the system to multiple servers across different domains and countries to attract a wider range of attackers. They also emphasize the need for search engine cooperation to enable real-time detection of emerging threats. Further research into automatically generating more accurate whitelists using tools like AutoRE is also proposed.

Summary updated

BibTeX
@inproceedings{john2011heat,
  author = {John P. John and Fang Yu and Yinglian Xie and Arvind Krishnamurthy and Martín Abadi},
  title = {{Heat-seeking honeypots: design and experience}},
  booktitle = {The Web Conference},
  year = {2011},
  doi = {10.1145/1963405.1963437},
}
Figure