Secure Training Methods for Federated Transfer Learning

The previous part introduced the model architecture of federated transfer learning; this part focuses on secure training methods required because gradient computation involves data from both parties, necessitating encryption algorithms, transmission protocols, and polynomial approximations.

Three combinations of encryption and protocol achieve secure training: additive homomorphic encryption, ABY framework, and SPDZ framework. Each method is described, including the use of ABY’s arithmetic sharing and SPDZ’s MAC‑based verification for malicious settings.

Polynomial approximation, specifically a second‑order Taylor expansion of the logistic function, is employed to convert non‑linear operations into linear ones that can be processed by the encryption schemes.

The article outlines a four‑step workflow for additive homomorphic encryption‑based federated transfer learning: (1) local data computation and initialization, (2) exchange of encrypted intermediate data, (3) computation of the other party’s encrypted gradients with added random masks, and (4) decryption and gradient update.

It also presents ABY‑ or SPDZ‑based methods that integrate the same polynomial approximation to securely compute gradients, noting that SPDZ can scale from two parties to n parties while tolerating dishonest majority.

Federated transfer learning is positioned alongside horizontal and vertical federated learning as one of the three main categories, with distinct characteristics in sample and feature overlap, and its implementation can be found in the open‑source FATE project.

The conclusion emphasizes the growing importance of privacy‑preserving model training and encourages readers to explore federated transfer learning techniques further.

References to relevant literature on homomorphic encryption, secure two‑party computation, and federated learning are provided.