Vertical FL Framework

This component performs privacy-preserving entity resolution to identify overlapping samples (e.g., shared customer IDs) across the vertically partitioned datasets of different parties. It is foundational, as VFL requires aligned entities to train a coherent model. Common implementations use cryptographic protocols like Private Set Intersection (PSI) to compute the intersection of entity sets without revealing non-overlapping entries. The efficiency and security of this engine directly impact the feasibility and trust in the entire VFL pipeline.

This module manages the split neural network architecture, defining the cut layer where the model is divided between the label owner and feature owners. It coordinates the distributed forward and backward passes (vertical forward/backpropagation). Key responsibilities include:

• Routing intermediate outputs and gradients between parties.
• Managing the lifecycle of model segments on different servers.
• Ensuring computational graphs remain consistent across the distributed system. It abstracts the complexity of the vertical training protocol, allowing developers to define model splits declaratively.

This suite of libraries provides the privacy-preserving primitives essential for VFL. It shields raw data by applying cryptographic techniques to the computations and communications within the framework. Core capabilities typically include:

• Homomorphic Encryption (HE): Enables computations (e.g., gradient aggregation) on encrypted intermediate results.
• Secure Multi-Party Computation (MPC): Protocols for joint computation where no party sees the others' private inputs.
• Differential Privacy (DP): Mechanisms to add calibrated noise to shared information (e.g., gradients) for formal privacy guarantees.
• Vertical Secure Aggregation: Specialized protocols for combining model updates from feature owners without leakage.

This component manages the significant vertical communication overhead inherent in VFL, where intermediate results must be exchanged in every training round. It optimizes for bandwidth, latency, and system heterogeneity by:

• Implementing compression techniques for gradients and embeddings.
• Scheduling asynchronous or staggered updates to handle stragglers.
• Batching communications to reduce round trips.
• Adapting to network conditions and participant availability. Its goal is to minimize the performance bottleneck caused by cross-party data transfer while maintaining protocol correctness.

This is the execution engine that runs the vertical training protocol and vertical inference protocol. It acts as the state machine for the collaborative process, handling:

• The iterative sequence of local computation, secure exchange, and aggregation.
• Fault tolerance and recovery for participant dropouts.
• Synchronization of model versions across all parties.
• The secure prediction process post-training, where feature owners contribute to an inference request without exposing their data. This runtime ensures the deterministic and reproducible execution of the VFL workflow.

This component provides observability into the decentralized training process, which is critical for enterprise adoption and algorithmic governance. It offers:

• Federated model evaluation metrics (e.g., global loss, accuracy) computed without centralizing test data.
• Tracking of contribution levels and data quality signals from each party.
• Auditing logs for all cryptographic operations and data accesses to ensure compliance with regulations.
• Tools to assess vertical model fairness and detect potential biases arising from the feature distribution across parties. It turns the "black box" of federated training into a transparent, managed system.

A dataset split where different features (columns) of the same samples (rows) are held by different parties. This is the foundational data structure for VFL.

• Example: Bank A holds customer credit histories (features), while E-commerce Platform B holds their purchase histories (different features) for the same set of users (aligned rows).
• This contrasts with horizontal partitioning, where different parties hold different rows with the same features.

The privacy-preserving process of identifying and matching the same real-world entities (e.g., customers, devices) across the vertically partitioned datasets held by different parties without revealing non-matching records.

• Purpose: Establishes the common set of samples on which the joint model will be trained.
• Primary Method: Uses cryptographic protocols like Private Set Intersection (PSI) to compute the overlap of entity IDs (e.g., user IDs, device IDs).

The model architecture used in VFL where the neural network is divided into multiple segments, each residing on a different party.

• Feature-Owner Segment: The initial layers that process local features. Computes an intermediate output.
• Label-Owner Segment: The subsequent layers, typically including the final output layer. Receives intermediate outputs to complete the forward pass and compute the loss.
• The division point is defined by the cut layer.

A cryptographic protocol that allows two or more parties, each holding a private set of items (e.g., user IDs), to compute the intersection of their sets without revealing any information about items not in the intersection.

• Critical Role in VFL: Used for the entity alignment phase before training begins.
• Ensures Privacy: Parties learn only which IDs they have in common, not the other party's full dataset.

The application of encryption schemes that allow computations on ciphertext. In VFL, this enables the label owner to perform operations on encrypted intermediate outputs from feature owners.

• Core Benefit: The label owner can compute gradients and update its model segment without ever decrypting the feature owners' data.
• Trade-off: Introduces significant computation overhead due to the complexity of encrypted arithmetic.

The defined sequence of communication and computation steps that coordinating parties follow to execute one training round in a VFL system.

• Typical Steps:
  1. Entity Alignment (via PSI).
  2. Vertical Forward Propagation: Feature owners compute to cut layer, send outputs (often encrypted) to label owner.
  3. Loss & Gradient Calculation: Label owner completes forward pass, computes loss and gradients for its segment.
  4. Vertical Backpropagation: Gradients are passed back securely to feature owners to update their segments.
• Governs the entire collaborative process.