Privacy-Friendly Collaboration for Cyber Threat Mitigation

Julien Freudiger, Emiliano De Cristofaro, Alex Brito

TL;DR

This paper introduces the "Sharing is Caring" (SIC) framework, a novel approach to collaborative cyber threat mitigation that allows organizations to privately estimate the benefits of data sharing and form coalitions without revealing sensitive information. Experiments on a massive real-world dataset show that this privacy-preserving collaboration can significantly improve predictive blacklisting accuracy by up to 105%.

What Problem Does This Paper Address?

Collaborative security, where organizations share threat intelligence, is widely recognized as a powerful way to enhance cyber defense. However, sharing sensitive security data (like firewall logs or attack patterns) poses significant challenges related to privacy, trust, and potential competitive disadvantage. Existing methods often involve sanitizing data (reducing its utility) or relying on central repositories (raising trust issues). This paper tackles the problem of enabling effective collaborative threat mitigation, specifically for predictive blacklisting (forecasting future attack sources), while rigorously preserving the privacy of participating organizations. The core challenge is to allow organizations to identify beneficial partners and share data in a way that maximizes prediction accuracy without compromising sensitive information.

Methods

The paper proposes the Sharing is Caring (SIC) framework, which comprises two main stages:

  1. Private Partner Selection: Organizations privately estimate the potential benefits of collaborating with each other. This is achieved using Secure Two-Party Computation (2PC) and related cryptographic primitives like Private Set Intersection Cardinality (PSI-CA) and Private Jaccard Similarity (PJS). Various metrics are used to quantify similarity between datasets, including:
  2. Set-based metrics: Intersection-Size and Jaccard Similarity (measuring overlap and relative overlap of unique attacking IPs).
  3. Correlation-based metrics: Pearson and Cosine similarity (capturing statistical relationships). Organizations then use these estimated benefits to decide whether to partner, often based on a threshold or by maximizing benefits.

  4. Privacy-Preserving Data Merging: Once partners are selected and form coalitions, they securely merge their datasets. The paper explores different merging strategies:

  5. Intersection with Associated Data: Partners share only attack events associated with IP addresses that attacked both parties. This is a conservative approach that reinforces knowledge about common adversaries.
  6. Union with Associated Data: Partners share all attack events from their combined datasets. This is a more generous approach that enriches datasets more rapidly. The paper emphasizes that the goal is to share only what is agreed upon, leveraging protocols like Private Set Intersection (PSI) with data transfer.

The core novelty lies in the combination of privacy-preserving techniques for both partner selection and data merging, enabling a distributed and trust-minimized collaborative security model.

Datasets

The primary dataset used for evaluation is from DShield.org. This dataset consists of approximately 2 billion suspicious IP address entries collected over two months. It includes contributor IDs, source IP addresses, target ports, and timestamps. The authors pre-processed this data, removing invalid entries, and focused on a subset of approximately 188,522 contributors to represent real-world scenarios with diverse attack patterns. This dataset is a real-world benchmark, differing from synthetic datasets by reflecting actual attack behaviors.

Key Insights and Ideas

  • Privacy-Preserving Collaboration is Feasible and Beneficial: The paper demonstrates that organizations can indeed collaborate to improve cyber threat detection without sacrificing privacy, challenging the notion that effective collaboration requires full data disclosure.
  • "Sharing is Caring" Framework: The SIC framework provides a structured, privacy-first approach to collaborative security, addressing both partner selection and data merging.
  • Intersection-Size is a Powerful Partner Selection Metric: The study highlights that simply looking at the number of common attacking IPs (Intersection-Size) is a highly effective metric for identifying beneficial partners, leading to significant accuracy improvements.
  • Sharing Common Attacks is Almost as Good as Sharing Everything: The finding that "Intersection with Associated Data" merging strategy performs nearly as well as "Union with Associated Data" is crucial. It suggests that organizations can achieve substantial predictive gains by sharing only the most relevant, overlapping threat information, thus enhancing privacy.
  • Collaboration Metrics Matter: The choice of metric for estimating collaboration benefits significantly impacts the outcome. Set-based metrics (like Intersection-Size) tend to favor larger, more active contributors, leading to greater accuracy improvements compared to correlation-based metrics.
  • Trust Can Be Established Automatically: The SIC framework's reliance on data similarity for partner selection implicitly builds trust, reducing the need for manual vetting processes.

Key Results

  • Up to 105% Prediction Improvement: Experiments on the DShield dataset showed that the proposed privacy-enhanced collaboration approach, particularly using the Intersection-Size metric for partner selection, achieved up to a 105% improvement in prediction accuracy for blacklisting.
  • Intersection-Size Outperforms Other Metrics: The Intersection-Size metric for partner selection consistently yielded the highest prediction improvements (mean of 105%), significantly outperforming Jaccard (mean of 58%) and correlation-based metrics like Pearson and Cosine (mean of ~40%).
  • Stable Coalitions: For most select metrics, coalitions of collaborators remained stable over time, with about 90% reuse of partners.
  • Low Operational Cost: The privacy-preserving protocols (PSI-CA, PSI) incurred relatively low computational and communication overhead, estimated to be around 40-55 seconds per entity for benefit estimation and merging, making the approach practical for daily use.

Limitations and Future Work

  • Dataset Bias: The authors acknowledge that the DShield dataset primarily consists of data from individuals and small organizations. This might bias the results, and it's unclear how well the findings generalize to larger enterprises.
  • Focus on True Positives (TP): The evaluation primarily focused on measuring True Positives (correctly predicted attacks) and did not consider false positives. While justified for prediction algorithms, real-world blacklisting systems need to balance both.
  • Semi-Honest Adversary Model: The security analysis assumes a semi-honest (honest-but-curious) adversary. Malicious adversaries who actively tamper with data are not considered.
  • Future Work: The authors suggest exploring other partner selection metrics (e.g., dissimilarity), adapting the approach to other collaborative security problems (spam filtering, malware detection, DDoS mitigation), and investigating incentive mechanisms. They also mention potential improvements like parallelization and sampling.

Summary updated

Abstract

Sharing of security data across organizational boundaries has often been advocated as a promising way to enhance cyber threat mitigation. However, collaborative security faces a number of important challenges, including privacy, trust, and liability concerns with the potential disclosure of sensitive data. In this paper, we focus on data sharing for predictive blacklisting, i.e., forecasting attack sources based on past attack information. We propose a novel privacy-enhanced data sharing approach in which organizations estimate collaboration benefits without disclosing their datasets, organize into coalitions of allied organizations, and securely share data within these coalitions. We study how different partner selection strategies affect prediction accuracy by experimenting on a real-world dataset of 2 billion IP addresses and observe up to a 105% prediction improvement.
BibTeX
@misc{freudiger2014privacy,
  author = {Julien Freudiger and Emiliano De Cristofaro and Alex Brito},
  title = {{Privacy-Friendly Collaboration for Cyber Threat Mitigation}},
  year = {2014},
  eprint = {1403.2123},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1403.2123},
}
Figure