SecretFlow Open‑Source Privacy Computing Framework Releases Version 0.7 with Enhanced MPC, Federated Learning, and Performance Optimizations

Open source has become a crucial model for technological innovation, and in the field of privacy computing, dozens of open‑source projects have emerged worldwide.

At the "Privacy Computing Open‑Source Forum" jointly organized by the Privacy Computing Alliance (PPCA) and the China Communications Standards Association Big Data Technical Standards Promotion Committee (CCSA TC601) on September 21, the SecretFlow trusted privacy‑computing open‑source framework was selected as one of the first projects to join the Alliance's open‑source working group.

Simultaneously, the first iteration of the SecretFlow framework (V0.7) was officially released, incorporating extensive community feedback and adding many features developers have requested.

The Alliance’s working group, established in July 2022, aims to promote excellent open‑source projects, improve the privacy‑computing ecosystem, and produce standards, reports, and project promotion activities.

SecretFlow’s architecture adopts a layered design with open capabilities at each level, abstracting both plaintext and ciphertext computation to lower development barriers and unify programming models for federated learning and secure multi‑party computation.

Version 0.7 introduces several key updates:

1. Multi‑party computation capabilities • Supports various 3‑party computation (3PC) machine‑learning algorithms such as LR, XGB, NN, as well as feature‑engineering algorithms like VIF, Binning, PearsonR, and 2‑party LR. • Adds PCG‑PSI for private set intersection, implementing the BC22 protocol, the most efficient current solution. • Provides DP‑PSI with differential‑privacy‑protected secure intersection, offering C++ source‑code kernels.

2. Federated learning capabilities • Adds logistic regression for mixed data partitioning scenarios. • Introduces new horizontal federated learning strategies with non‑IID optimization and communication optimizations. • Enhances horizontal and vertical federated DNNs with differential‑privacy protection.

3. Infrastructure and performance optimizations • Introduces HEU.numpy module for higher‑performance encrypted‑plaintext matrix operations. • Optimizes SPU ABY3 performance and memory usage; Cheetah matrix multiplication gains up to 5× speedup.

4. Documentation improvements • Provides Chinese‑language documentation. • Refines documentation structure for smoother user experience. • Enriches algorithm and architecture design explanations.

SecretFlow will continue quarterly high‑frequency iterations, strengthen community interaction, and expand the SecretFlow Open Platform to meet diverse development needs, accelerating the practical deployment of privacy‑computing technologies across AI and data‑analysis scenarios.