Shareprom: A Tool for Privacy-Preserving Inter-Organizational Process Mining

Gamal Elkoumy, Stephan A. Fahrenkrog-Petersen, Marlon Dumas, Peeter Laud, Alisa Pankova, Matthias WeidlichBPM (PhD/Demos) 2020 · OpenReview

Introduction

Process mining techniques are essential for discovering and optimizing business processes by extracting insights from event logs. However, these techniques typically operate within a single trust domain. In modern collaborative environments, such as global supply chains, processes span multiple organizations. Sharing raw event logs across these boundaries is often prohibited by confidentiality concerns and strict data protection regulations like GDPR.

Shareprom addresses this challenge by enabling joint process mining without requiring organizations to share their raw data. By integrating Secure Multi-Party Computation (MPC) and Differential Privacy (DP), Shareprom allows parties to generate a shared Directly-Follows Graph (DFG) while maintaining strict data confidentiality.

Methodology

The Shareprom architecture is built upon two core privacy-enhancing pillars: Secure Multi-Party Computation and Differential Privacy.

System architecture of Shareprom showing the secure multi-party computation workflow.

Secure Multi-Party Computation (MPC)

Shareprom utilizes the Sharemind platform to transform local event data into "secret shares." These shares are distributed across three independent computing parties. No single party can reconstruct the original data; instead, they perform mathematical operations directly on the shares. This protocol is designed to be secure against "honest-but-curious" adversaries.

Differential Privacy (DP)

To prevent information leakage from the final output, Shareprom applies Differential Privacy. Even if the computation itself is secure, the resulting process map could potentially reveal sensitive details, such as specific transition times. Shareprom injects Laplacian noise into the final DFG frequencies and time annotations, providing a mathematical guarantee that individual data points cannot be reverse-engineered.

The Workflow

  1. Preprocessing: Organizations pad their logs to ensure uniform trace lengths, preventing metadata leakage based on trace size.
  2. Secret Sharing: Data is uploaded to Sharemind servers as randomized shares.
  3. MPC Calculation: Servers compute the DFG matrix (frequencies and execution times) in the encrypted domain.
  4. Noisy Release: Differential privacy noise is added before the final process map is released to the analyst.

Key Insights

A critical insight in Shareprom's design is that data volume and trace length can act as side channels. By enforcing uniform trace lengths through client-side padding, the system mitigates leaks based on the structural characteristics of the logs. Furthermore, the three-party model—two data providers and one analysis firm—aligns with common industrial consulting and auditing workflows. By focusing on the DFG, Shareprom remains compatible with industry-standard process mining tools.

Results and Evaluation

Shareprom successfully generates time-annotated process maps from distributed logs without intermediate data declassification. The system has been evaluated using logs containing thousands of events and dozens of distinct activity labels, demonstrating its feasibility for mid-sized organizational processes.

Shareprom user interfaces for organizations and analysis firms, alongside the resulting joint Directly-Follows Graph (DFG).

The authors developed a Python-based desktop client that facilitates the import of standard XES files. This interface allows users to interact with the MPC servers seamlessly, as shown in the figure above. Performance benchmarks indicate the tool is scalable for many practical business scenarios.

Limitations and Future Work

While Shareprom proves the viability of privacy-preserving inter-organizational process mining, several areas for improvement remain:

  • Scalability: Future optimizations are required to handle "Big Data" logs with millions of events or hundreds of unique activity types.
  • Analytical Scope: The current implementation is limited to DFGs. Expanding to conformance checking and advanced discovery algorithms is a priority.
  • Utility-Privacy Trade-off: Further research is needed to optimize the balance between the level of DP noise and the practical utility of the resulting process maps for business analysts.

Summary updated

BibTeX
@inproceedings{elkoumy2020shareprom,
  author = {Gamal Elkoumy and Stephan A. Fahrenkrog-Petersen and Marlon Dumas and Peeter Laud and Alisa Pankova and Matthias Weidlich},
  title = {{Shareprom: A Tool for Privacy-Preserving Inter-Organizational Process Mining}},
  booktitle = {BPM (PhD/Demos)},
  year = {2020},
  url = {https://openreview.net/forum?id=fgVR7YLcNW},
}
Figure