honest-but-curious

The term "honest-but-curious" is a security model often considered in the context of Federated Learning (FL) and cryptographic protocols like Secure Multi-Party Computation (SMPC). In this model, participants are assumed to follow the protocol rules. Still, they are also curious to learn additional information about other participants' private data if they can do so within the confines of those rules.

Characteristics of Honest-but-Curious Participants

1. Protocol Adherence. They execute the protocol as specified without deviation. This means they won't intentionally corrupt the computation or disseminate false information.
2. Curiosity. Although they stick to the protocol, they are curious about learning any additional information that might be gleaned from the data exposed during the protocol's execution.

Implications for Federated Learning

In Federated Learning, multiple parties collaborate to build a machine-learning model without sharing their raw data. While the focus is to create a collective model that benefits all, each party is also interested in protecting sensitive data.

1. Data Privacy. In an honest-but-curious setting, parties would not break the protocol to steal data. However, they may use data exposed during honest protocol execution to learn more about others.
2. Model Inference. Even if a participant follows the Federated Learning protocol correctly, they may still attempt to perform model or data inference attacks to understand something about other participants' data based on the model updates they receive.