Sherpa.ai Federated Learning Platform

WIDE RANGE OF
AGGREGATORS

There are a multitude of aggregation processes, but not all of them are suitable for all models or paradigms neither offer security in untrusted environments or achieve the best accuracy in certain scenarios. A platform must offer a wide range of aggregators with which to approach learning processes of the most diverse nature and ensure that the best result is obtained from the existing model and available data.
Specific aggregators have been designed for Vertical Federated Learning (VFL). One of them is the concatenation aggregator, which, on the one hand, allows to unlock the potential of heterogeneous data; on the pother hand, assisted by Differential Privacy (DP) or Secure Multi-Party Computation (SMPC) preserves data privacy.

The federated architecture itself naturally protect data privacy. However, Adversarial Attacks may be performed to retrieve some information anyhow. Robust aggregators are designed to defend against such attacks, thus enforcing the privacy-resiliency of the federated architecture.

ROBUST
AGGREGATORS

Distributed environments architectures for Federated Learning (FL) add vulnerabilities, such as the ability to defend against adversary attacks. In classical machine learning, the majority of adversary attacks can be prevented through data sanitation techniques, where the data is inspected by cleaning it of possible contaminated data. Clearly the constraint that training data never leaves the clients in FL makes it impossible to apply these defense mechanisms and leaves a small margin of maneuver. For this reason, the development of ad-hoc defense mechanisms for FL is crucial.

Robust aggregators provide defenses against poisoning attacks (backdoor and byzantine), avoiding sample missclassification, label flipping, the introduction of samples out of the training distribution and random model updates.