Skip to main content
. 2022 May 3;24(5):643. doi: 10.3390/e24050643

Figure 1.

Figure 1

Threat model. The user sends the private representations to the server for final classification. Eavesdropping on the private features, the adversary wishes to reconstruct the original data or infer sensitive attributes. The adversary does not have access to the local obfuscation mechanism used by the user, shown in blue dashed lines.