Federated Learning vs Secure Multi‑Party Computation: Concepts, Challenges, and Alibaba’s Solutions

Federated learning (FL) was introduced by Google in 2016 to enable distributed model training on Android devices without collecting raw user data, using a gradient‑sharing approach that preserves privacy by design. The article describes the evolution of FL, its architecture (parameter server), and the distinction between Google’s cross‑device FL and China’s cross‑silo FL, which typically involves a small number of institutions collaborating on vertically partitioned data for credit or advertising use cases.

The security challenges of FL are examined, including the risk of reconstructing raw data from gradients, the limitations of differential privacy (accuracy loss) and secure aggregation (requires many participants), and the vulnerabilities that arise when only a few parties are involved. The article also highlights the difficulties of vertical FL, such as sample alignment, handling unlabeled parties, and computing weight‑of‑evidence (WOE) without leaking label information.

Secure multi‑party computation (MPC) is introduced as a cryptographic technique that enables parties to jointly compute a function while revealing nothing beyond the final result. The text outlines secret‑sharing based MPC, illustrating how parties split their data into random shares, exchange them, and perform arithmetic operations (addition, multiplication, division, comparison) on the shares to compute models such as logistic regression without exposing intermediate values.

Practical examples show how MPC can be used to compute WOE securely: the label‑holding party provides a secret‑shared vector of positive/negative samples, the other party multiplies it with its own secret‑shared feature counts, and then performs secret‑shared division and logarithm to obtain the WOE value. The article also notes that MPC eliminates the need for data alignment, as secret‑shared matching can be performed without revealing user identities, offering compliance advantages under regulations like GDPR.

Finally, the article discusses the performance trade‑offs of MPC versus FL, acknowledging that MPC incurs higher communication overhead per operation, which can limit scalability for complex models (e.g., XGBoost). Nevertheless, recent optimizations have made MPC practical for large‑scale logistic regression, as demonstrated by Alibaba’s championship in the iDASH competition, and the open‑source FATE framework now supports the SPDZ protocol for further development.