Detecting Targeted Attacks By Multilayer Deception

Wei Wang, Jeffrey Bickford, Ilona Murynets, Ramesh Subbaraman, Gokul SingarajuJournal of Cyber Security and Mobility 2013 · DOI

TL;DR

This paper introduces a novel multi-layer deception system designed to detect sophisticated, multi-stage targeted attacks that are notoriously difficult to identify with traditional methods. By strategically placing "honey" entities across different layers of an enterprise network, the system aims to significantly increase the probability of detecting an attacker early in their intrusion process. Furthermore, the authors propose an optimization framework to determine the most cost-effective deployment of these deception mechanisms, minimizing overall expected losses.

What Problem Does This Paper Address?

Modern enterprises are increasingly targeted by highly customized and sophisticated attacks. Unlike broad, opportunistic attacks, these targeted assaults are executed by skilled adversaries who meticulously plan their intrusions, often employing multi-stage strategies. These attackers aim to steal sensitive data or intellectual property and are adept at leaving minimal traces, making detection a significant challenge. Traditional security measures, such as signature-based detection or blacklists, are often ineffective against these novel and evolving threats. The paper addresses the critical need for a more proactive and layered defense strategy that can identify and thwart these advanced persistent threats before significant damage occurs.

Methods

The core of the proposed solution is a multi-layer deception system. This system operates by creating "honey" entities that mimic real assets within an enterprise network. These honey entities are strategically placed to intercept attackers at different stages of their intrusion. The paper outlines three primary layers of deception:

  • Honey People (HP): This layer aims to protect against social engineering attacks, particularly spear-phishing. It involves creating fake employee profiles (e.g., with alternative email addresses) that can be posted on public platforms. When an attacker targets these fake profiles, an alert is triggered.
  • Honey Files with Honey Activity (HFHA): This layer focuses on detecting attackers who have already gained access to a user's local machine. It involves creating deceptive files and folders that are indistinguishable from real ones. Crucially, "honey activity" is incorporated to simulate user interactions with these files, making them appear more realistic and harder for attackers to ignore or bypass.
  • Honey Servers with Honey Activity (HSHA): This layer targets attackers who are attempting to access global assets (like databases). It involves deploying deceptive servers that mirror legitimate ones. Similar to HFHA, "honey activity" is used to emulate user network behavior, such as connecting to these servers, to detect malicious access patterns.

The system integrates these layers to provide a comprehensive defense. Alerts from these deception entities are sent to an Analyst Server for correlation and analysis.

Beyond the deception mechanisms, the paper introduces an optimization model. This model formulates the problem of deploying deception entities as an optimization task. The goal is to minimize the enterprise's total expected loss (combining system deployment costs and potential asset compromise losses) while adhering to a defined budget. This involves calculating probabilities of compromise through direct and indirect social connections and then determining the optimal allocation of honey people and honey files.

Datasets

Not discussed in the paper. The paper focuses on a conceptual framework and an optimization model, with a simulated example for evaluation rather than real-world datasets.

Key Insights and Ideas

  • Layered Deception is Key: The paper strongly advocates for a multi-layered approach to deception, arguing that each layer significantly enhances the probability of early detection. This is crucial for combating multi-stage attacks where attackers progress through distinct phases.
  • Deception as Proactive Defense: Instead of solely relying on detecting known malicious patterns, the system proactively lures attackers into revealing themselves by interacting with fake assets.
  • The "Honey Activity" Advantage: Simply creating fake files or servers isn't enough. The "honey activity" component is vital for making these deceptive entities appear as part of normal user behavior, thus tricking sophisticated attackers who might otherwise ignore static decoys.
  • Optimization for Cost-Effectiveness: The introduction of an optimization model to balance the cost of deploying deception entities against the potential losses is a significant contribution. It highlights that more isn't always better, and there's an optimal point for investment.
  • Over-Deployment Can Be Detrimental: A surprising insight is that exceeding the optimal budget for deception entities can actually increase total expected expenses, as the cost of deployment outweighs the marginal reduction in expected loss.

Key Results

  • Enhanced Detection Probability: The multi-layer deception system is expected to greatly enhance the probability of detecting sophisticated attacks at an early stage.
  • Optimal Resource Allocation: The optimization model provides a method to find the most effective allocation of deception resources (honey people and honey files) given a budget, minimizing total expected loss.
  • Example Scenario: In a simulated example with ten employees, a budget of $500,000 for honey entities reduced the expected loss from $122 million to $10 million.
  • Budget Optimization: The study found an optimal budget of $1.8 million that minimized total expected expenses (deployment cost + expected loss). Exceeding this budget led to increased total expenses.

Limitations and Future Work

  • Proof of Concept: The paper presents a proof-of-concept implementation for protecting local assets, indicating that the full multi-layer system is still under development.
  • Manual Processes: The current implementation of honey file generation is manual, which the authors aim to automate.
  • Administrator Privileges: The deception service runs with administrator privileges, making it vulnerable to malware that also gains administrator rights. A hypervisor-based approach is suggested for higher security.
  • False Positives: Reducing false positives is acknowledged as a challenging task. The paper suggests employee awareness and technical differentiators (like secondary channels for identifying real files) as mitigation strategies.
  • Cloud Environment: Future work includes implementing the multi-layer deception approach in cloud environments, leveraging Virtual Desktop Infrastructure (VDI) for more centralized management of honey files.
  • Cost and Value Estimation: Further research is needed to better estimate the cost of deploying honey entities and the value of each asset, potentially through surveys and real-world case studies.
  • Interactions and Assumptions: The paper mentions studying interactions between honey activities and honey servers, and exploring more scenarios with different assumptions on targeted assets.

Summary updated

Abstract

Over the past few years, enterprises are facing a growing number of highly customized and targeted attacks that use sophisticated techniques and seek after important company assets, such as customer data and intellectual property. Unlike conventional attacks, targeted attacks are operated by experts who use multiple steps to gain access to sensitive assets, and most of time, leave very few network traces behind for detection. In this paper, we propose a multi-layer deception system that provides an in depth defense against such sophisticated targeted attacks. Specifically, based on previous knowledge and patterns of such attacks, we model the attacker as trying to compromising an enterprise network via multiple stages of penetration and propose defenses at each of these layers using deception based detection. Due to multiple layers of deception, the probability of detecting such an attack will be greatly enhanced. We present a proof of concept implementation of one of the key deception methods proposed. Due to various financial constraints of an enterprise, we also model the design of the deception system as an optimization problem in order to minimize the total expected loss due to system deployment and asset compromise.We find that there is an optimal solution to deploy deception entities,and even over spending budget on more entities will only increase the total expected loss to the enterprise. Such a system can be coupled with existing detection techniques to protect enterprises from sophisticated attacks.
BibTeX
@article{wang2013detecting,
  author = {Wei Wang and Jeffrey Bickford and Ilona Murynets and Ramesh Subbaraman and Gokul Singaraju},
  title = {{Detecting Targeted Attacks By Multilayer Deception}},
  journal = {Journal of Cyber Security and Mobility},
  year = {2013},
  doi = {10.13052/jcsm2245-1439.224},
}
Figure