Skip to main content

Inference

The risk of inference refers to the phenomenon where seemingly anonymous or harmless information can be used to deduce sensitive attributes about an individual. Unlike individualization, where the goal is to uniquely identify an individual, inference involves drawing conclusions from a subset of data without needing to directly know the person’s identity. A dataset may be free from individualization for a set of attributes, yet still be vulnerable to inference attacks.

Example

Imagine you learn that your 25-year-old neighbor, who is 1.80 meters tall, recently participated in a clinical study. You access a database containing information about participants in this study, and find three individuals who match these characteristics. While there is no individualization risk – as you cannot know which of the three is your neighbor – you notice that all three participants are sick. You can then infer that your neighbor is probably sick because the data suggests a strong correlation between these characteristics and the illness.

List of metrics measuring singling out criteria​

Factors Influencing the Risk of Inference​

The risk of inference is directly related to the density of the data and the precision of the information collected. Even if a dataset does not directly identify an individual, it can still allow for additional information about that individual to be inferred, enriching an attacker’s knowledge base.

Importance of Mitigation​

To limit the risks of inference, it is crucial to understand that even a dataset that appears anonymous at first glance can provide enough information to allow inferences about individuals. The metrics developed in our system evaluate not only the risk of individualization but also the risks of inference, helping to better protect sensitive data. Mitigating this risk requires considering interactions between different variables and adopting a more sophisticated approach to ensure that no sensitive conclusions can be drawn from the available data.

It’s important to note that for inference to be considered present, the inferred information must be certain. For instance, determining that an individual known to be a smoker has a 90% chance of contracting lung cancer would not be considered inference.