Skip to content

Latest commit

 

History

History
110 lines (93 loc) · 7.8 KB

RELEASE.md

File metadata and controls

110 lines (93 loc) · 7.8 KB
Raw

Release 2.2.0

Major Features and Improvements

Deployment

  • Upgrade from Python 3.8 to Python 3.10
  • Upgrade the PyTorch version to 2.x

Release 2.1.1

Major Features and Improvements

Component

  • Support server model saving in Homo-NN

ML

  • aggregator support aggregation of torch.bfloat16 data type

Release 2.1.0

Major Features and Improvements

Arch

  • Some bugs fixed for spark computing engine

Component

  • Unified IO keys naming format for all components
  • Add LLMLoader to support running FATE-LLM v2.0 with pipeline

OSX

  • Compatible with eggroll-v2.x

Release 2.0.0

Major Features and Improvements

Arch 2.0：Building Unified and Standardized API for Heterogeneous Computing Engines Interconnection

  • Introduce Context to manage useful APIs for developers, such as Distributed Compting, Federation, Cipher, Tensor, Metrics, and IO.
  • Introduce Tensor data structure to handle local and distributed matrix operation, with built-in heterogeneous acceleration support.
    • abstracted PHETensor, smooth switch between various underlying PHE implementations through standard interface
  • Introduce DataFrame, a 2D tabular data structure for data io and simple feature engineering
    • add data block manager to support mixed-type columns & feature anonymization
    • added 30+ operator interfaces for statistics, including comparison, indexing, data binning, and transformation, etc
  • Refactor Federation, a unified interface for federated computing. We provide a unified Serdes control and more user-friendly api.
  • Introduce Config, a unified configuration for FATE, including safety restrictions, system configuration, and algorithm configuration
  • Refactor logger, customizable logging for different use cases and flavors.
  • Introduce Launcher, a simple tool for federated program execution, especially useful for standalone and local debugging
  • Framework: PSI-ECDH protocol support, single entry for histogram statistical computation
  • Deepspeed integration: support distributed training using deepspeed with Eggroll.
  • Protocol: Support for SSHE(mpc and homomophic encryption mixed protocol), ECDH, Secure Aggregation protocols
  • Experimental Integrate Crypten for SMPC support, more protocols and features will be added in the future

Components 2.0: Building Standardized Algorithm Components for different Scheduling Engines

  • Introduce components toolbox to wrap ML modules as standard executable programs
  • spec and loader expose clear API for smooth internal extension and external system integration
  • Provide several cli tools to interact and execute components
  • Input-Output: Further decoupling of FATE-Flow, providing standardized black-box calling processes
  • Component Definition: Support for typing-based definition, automatic checking for component parameters, support for multiple types of data and model input and output, in addition to multiple inputs

ML 2.0: Major functionality migration from FATE-v1.x, decoupling call hierarchy

  • Data preprocessing: Added DataFrame Transformer; Reader, Union and DataSplit migration completed
  • Feature Engineering: Migrated HeteroFederatedBinning, HeteroFeatureSelection, DataStatistics, Sampling, FeatureScale and Pearson Correlation
  • Federated Training Migrated: HeteroSecureBoost, HomoNN, HeteroCoordinatedLogisticRegression, HeteroCoordinatedLinearRegression, SSHE-LogisticRegression and SSHE-LinearRegression
  • Federated Training Added:
    • SSHE-HeteroNN: based on mpc and homomorphic encryption mixed protocal
    • FedPASS-HeteroNN: based on fedpass protocol

Algorithm Performance Improvements: (Comparison with FATE-v1.11.*)

  • PSI (Privacy Set Intersection): tested on a dataset of 100 million with an intersection result of 100 million, 1.8+ times of FATE-v1.11.4
  • Hetero-SSHE-LR: tested on data of guest 10w * 30 dimensions and host 10w * 300 dimensions, 4.3+ times of FATE-v1.11.4
  • Hetero-NN(Based on FedPass Protocol): tested on data of guest 10w * 30 dimensions and host 10w * 300 dimensions, basically consistent with the plaintext performance, 143+ times of FATE-v1.11.4
  • Hetero-Coordinated-LR: tested on data of guest 10w * 30 dimensions and host 10w * 300 dimensions, 1.2+ times of FATE-v1.11.4
  • Hetero-Feature-Binning: tested on data of guest 10w * 30 dimensions and host 10w * 300 dimensions, 1.5+ times of FATE-v1.11.4

OSX(Open Site Exchange) 1.0: Building Open Platform for Cross-Site Communication Interconnection

  • Implement the transmission interface in accordance with the “ Technical Specification for Financial Industry Privacy Computing Interconnection Platform”,The transmission interface is compatible with FATE 1.X version and FATE 2.X version
  • Supports GRPC synchronous and streaming transmission, supports TLS secure transmission protocol, and is compatible with FATE1.X rollsite components
  • Supports Http 1.X protocol transmission and TLS secure transmission protocol
  • Support message queue mode transmission, used to replace rabbitmq and pulsar components in FATE 1.X
  • Supports Eggroll and Spark computing engines
  • Supports networking as an Exchange component, with support for FATE 1.X and FATE 2.X access
  • Compared to the rollsite component, it improves the exception handling logic during transmission and provides more accurate log output for quickly locating exceptions.
  • The routing configuration is basically consistent with the original rollsite, reducing the difficulty of porting
  • Supports HTTP interface modification of routing tables and provides simple permission verification
  • Improved network connection management logic, reduced connection leakage risk, and improved transmission efficiency
  • Using different ports to handle access requests both inside and outside the cluster, facilitating the adoption of different security policies for different ports

FATE Flow 2.0: Building Open and Standardized Scheduling Platform for Scheduling Interconnection

  • Adapted to new scalable and standardized federated DSL IR
  • Built an interconnected scheduling layer framework, supported the BFIA protocol
  • Optimized process scheduling, with scheduling separated and customizable, and added priority scheduling
  • Optimized algorithm component scheduling，support container-level algorithm loading， enhancing support for cross-platform heterogeneous scenarios
  • Optimized multi-version algorithm component registration, supporting registration for mode of components
  • Federated DSL IR extension enhancement: supports multi-party asymmetric scheduling
  • Optimized client authentication logic, supporting permission management for multiple clients
  • Optimized RESTful interface, making parameter fields and types, return fields, and status codes clearer
  • Added OFX(Open Flow Exchange) module: encapsulated scheduling client to allow cross-platform scheduling
  • Supported the new communication engine OSX, while remaining compatible with all engines from FATE Flow 1.x
  • Decoupled the System Layer and the Algorithm Layer, with system configuration moved from the FATE repository to the Flow repository
  • Published FATE Flow package to PyPI and added service-level CLI for service management
  • Migrated major functionality from FATE Flow 1.x

Fate-Client 2.0: Building Scalable Federated DSL for Application Layer Interconnection And Providing Tools For Fast Federated Modeling.

  • Introduce new scalable and standardized federated DSL IR(Intermediate Representation) for federated modeling job
  • Compile python client to DSL IR
  • Federated DSL IR extension enhancement: supports multi-party asymmetric scheduling
  • Support mutual translation between Standardized Fate-2.0.0 DSL IR and BFIA protocol
  • Support using components with BFIA protocol through adapter mode
  • Migrated Flow CLI and Flow SDK

Fate-Test: FATE Automated Testing Tool

  • Migrated automated testing for functionality, performance, and correctness

Easy Deploy

  • Supports installation of FATE by PyPi