MixNN: Protection of Federated Learning Against Inference Attacks by Mixing Neural Network Layers

TL;DR

This paper presents MixNN a proxy-based privacy-preserving system for federated learning to protect the privacy of participants against a curious or malicious aggregation server trying to infer sensitive information (i.e., membership and attribute inferences).

Abstract

Machine Learning (ML) has emerged as a core technology to provide learning models to perform complex tasks. Boosted by Machine Learning as a Service (MLaaS), the number of applications relying on ML capabilities is ever increasing. However, ML models are the source of different privacy violations through passive or active attacks from different entities. In this paper, we present MixNN a proxy-based privacy-preserving system for federated learning to protect the privacy of participants against a curious or malicious aggregation server trying to infer sensitive attributes. MixNN receives the model updates from participants and mixes layers between participants before sending the mixed updates to the aggregation server. This mixing strategy drastically reduces privacy without any trade-off with utility. Indeed, mixing the updates of the model has no impact on the result of the aggregation of the updates computed by the server. We experimentally evaluate MixNN and design a new attribute inference attack, Sim, exploiting the privacy vulnerability of SGD algorithm to quantify privacy leakage in different settings (i.e., the aggregation server can conduct a passive or an active attack). We show that MixNN significantly limits the attribute inference compared to a baseline using noisy gradient (well known to damage the utility) while keeping the same level of utility as classic federated learning.

Figures (9)

• Figure 1: Example of Neural Network.
• Figure 2: Operating flow of Federated Learning.
• Figure 3: MixNN introduces a proxy which receives the parameter updates from each participant, shuffle them to remove attribute footprint before to route them to the aggregation server.
• Figure 4: $\nabla$Sim infers attributes according to the gradient vector returned by participants (i.e., the parameter updates) and the learning models representative to each class of sensitive attributes (background knowledge).
• Figure 5: MixNN provides the same utility than a standard FL scheme, noisy gradient however decreases significantly the utility and slows down the convergence.