Security Analysis of Federated Learning and Split Learning
Transferability of attacks and defenses between federated learning and split learning.
University of Maryland - Baltimore County, USA · Supervisor: Dr. Md Tariqul Islam · January 2026 – Present
Overview
Federated learning and split learning both train models without centralizing raw data, but they distribute computation differently — federated learning keeps the full model on every client, while split learning cuts the network and shares intermediate activations instead. Attacks and defenses have largely been developed within each paradigm separately, and it is not established which results carry across. This work asks where the architectural difference invalidates the assumptions an attack or a defense depends on.
Approach
1
Systematize the landscape
Catalog attack patterns and defense mechanisms across both paradigms, identifying the threat model each was designed against.
2
Test generalizability
Evaluate model inversion, membership inference, and gradient leakage against differential privacy and secure aggregation in both settings, measuring where effectiveness holds and where it collapses.
3
Establish transferability criteria
Define the theoretical grounds and evaluation metrics for when a result in one paradigm should be expected to hold in the other.
Status
A Systematization of Knowledge (SoK) paper is in preparation, targeting the IEEE Symposium on Security and Privacy (Oakland) 2027.
Tools
Python PyTorch FedAvg FedProx Differential Privacy Secure Aggregation