Study of Privacy in Federated Reservoir Computing
Christopher C. Zawacki, Eyad H. Abed · 2024
Federated Reservoir Computing allows multiple parties to collaboratively train a reservoir-based model while providing privacy guarantees for participating clients. One such proposed method is Incremental Federated Learning [1]. This method trains Echo State Networks (ESN) within a federated framework by exploiting an algebraic decomposition of the objective function. Incremental Federated Learning relies on the complexity of the reservoir to ensure the privacy of client data. In this work, we present an example showing that, in actuality, client data may be reconstructed by an adversarial agent with access to the suggested update step in Incremental Federated Learning. We demonstrate that client input data can be extracted from the transferred matrices, suggesting the need for additional privacy safeguards when extending Reservoir Computing to federated environments.