Abstract
Event logs recorded during the execution of business processes constitute a valuable source of information. Applying process mining techniques to them, event logs may reveal the actual process execution and enable reasoning on quantitative or qualitative process properties. However, event logs often contain sensitive information that could be related to individual process stakeholders through background information and cross-correlation. We therefore argue that, when publishing event logs, the risk of such re-identification attacks must be considered. In this paper, we show how to quantify the re-identification risk with measures for the individual uniqueness in event logs. We also report on a large-scale study that explored the individual uniqueness in a collection of publicly available event logs. Our results suggest that potentially up to all of the cases in an event log may be re-identified, which highlights the importance of privacy-preserving techniques in process mining.
The Privacy Challenge in Process Mining
Process mining is a powerful field that analyzes digital footprints—event logs—to improve business efficiency. However, these logs often contain sensitive data about patients, employees, or customers. While researchers typically remove direct identifiers like names, the remaining data—timestamps, activity sequences, and case attributes—can act as a unique "fingerprint."
The primary challenge is the lack of a standardized framework to quantify the risk a specific event log carries before publication. Data owners are often hesitant to share logs due to vague privacy concerns, while researchers may inadvertently publish logs that are easily de-anonymized. This research provides a formal framework to measure this re-identification risk through the lens of "individual uniqueness."
Measuring Individual Uniqueness
The study proposes two primary measures for quantifying uniqueness:
- Uniqueness based on Case Attributes: This focuses on static information tied to a process instance (e.g., a patient's age and sex). A statistical model (Gaussian copulas) is used to estimate "population uniqueness"—the probability that an individual is unique not just within the sample log, but within the entire underlying population.
- Uniqueness based on Traces: This focuses on the sequence of events, defining a trace as a series of points (activity, timestamp, and event attributes). It measures how many random "points" an adversary needs to know to uniquely identify a specific trace among all others in the log.
To test these measures, the authors utilize Projections—subsets of data (e.g., only activities, or activities + timestamps)—to evaluate how different levels of data disclosure affect privacy and utility.
Experimental Setup and Datasets
The study evaluated 12 publicly available event logs from the 4TU.Centre for Research Data. These were categorized into:
- Real-life individuals (R): Logs where cases refer to natural persons (e.g., the Sepsis healthcare log, BPI Challenge logs, and the ADL daily living log).
- Software (S): Logs referring to technical components (e.g., Apache, JUnit, NASA).
The evaluation primarily focused on the "Real-life individuals" category, with detailed deep-dives into the BPI Challenge 2018 (agricultural payments) and Sepsis (hospital emergency room) logs.
Key Insights and Results
- The Power of Timestamps: Timestamps are incredibly identifying. In the Sepsis log, knowing just four random events (the activity name and the exact time it occurred) was enough to uniquely identify 100% of the traces.
- Traces vs. Attributes: While static case attributes (like age or department) can be identifying, the behavioral data in the traces (the "how" and "when" of a process) is significantly more unique. Across 11 of the 12 evaluated logs, the uniqueness was greater than 62%, even when the adversary only had a random selection of trace points.
- Generalization Limits: Reducing timestamp resolution (e.g., changing exact times to dates) reduces uniqueness but often leaves up to 70% of the population identifiable if the trace is long enough.
- BPI Challenge 2018 Findings: Using all case attributes, the estimated population uniqueness was approximately 97%. Even with a single attribute like payment type, 16% of the population remained unique.
Conclusion and Future Work
This research demonstrates that "anonymized" event logs are highly susceptible to re-identification. By quantifying the "privacy-utility trade-off," the paper highlights that while data minimization (removing event-level metadata) drops uniqueness sharply, it may also limit the depth of process mining analysis.
Future work is needed to develop optimized mitigation strategies, such as differential privacy or k-anonymity, to protect individual privacy without sacrificing the analytical value of event logs. The study assumes an adversary selects trace points randomly; in reality, an attacker knowing rare events would find re-identification even easier, suggesting the risks reported here may be conservative.
Summary updated