Distributed Model Training With Scalar Gradient Projections

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Federated Learning faces challenges such as security against Byzantine clients and communication efficiency, particularly in scenarios where edge devices have limited computation and communication capabilities.

Innovation Solution

A method involving a central node and client nodes that uses a shared seed for pseudo-random number generation to compute a scalar reflecting the gradient projection, allowing clients to send only scalars instead of high-dimensional vectors, coupled with robust aggregation techniques to enhance privacy and security.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If clients send high-dimensional gradient vectors to the central node, then model training accuracy is maintained, but communication bandwidth consumption increases significantly

Engineering Contradiction:
Improvemodel training accuracyVSAvoidcommunication bandwidth consumption
Core Design Contradiction:
Measurement precisionVSQuantity of substance

Solution Approach 1:

The patent extracts only the essential information from the full gradient vector by computing a scalar projection along a random direction. Instead of transmitting the complete high-dimensional gradient vector, each client sends only a single scalar value representing the gradient's projection onto a randomly generated direction vector. This extraction principle maintains the critical directional information needed for model updates while dramatically reducing communication overhead from O(d) to O(1) per client per round.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent transforms the gradient communication paradigm by changing the parameter representation from full-dimensional vectors to scalar projections. By projecting gradients onto random directions and transmitting only these scalar values, the system fundamentally alters the communication parameter format. This parameter transformation enables efficient aggregation at the central node while preserving the essential gradient information needed for model convergence.

Inventive Principle:
Principle #35Parameter changes

2Productivity

If the system aggregates updates from all client nodes, then model convergence is improved, but vulnerability to Byzantine attacks increases

Engineering Contradiction:
Improvemodel convergence speedVSAvoidsecurity against Byzantine clients
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The patent introduces random direction vectors as intermediaries between clients and the central node. These random directions serve as a mediating layer that obscures the actual gradient information while preserving its essential directional properties. By projecting gradients onto these intermediary random vectors before aggregation, the system maintains convergence properties while making it difficult for Byzantine clients to craft targeted attacks, as the random directions mask the true gradient structure.

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent applies partial action by transmitting only a scalar projection rather than the complete gradient vector. This partial information transmission is sufficient for model convergence but insufficient for Byzantine clients to execute effective attacks. The scalar projection contains just enough directional information to guide model updates while lacking the detailed structure needed for sophisticated adversarial manipulation.

Inventive Principle:
Principle #16Partial or excessive action

3Measurement precision

If clients perform backpropagation on local data, then training accuracy is maintained, but computation and memory usage on edge devices increases

Engineering Contradiction:
Improvetraining accuracyVSAvoidcomputation and memory requirements
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent extracts the essential gradient information needed for model updates by computing only a scalar projection rather than performing complete backpropagation. Clients compute the gradient projection onto a random direction and transmit only this scalar value to the central node. This extraction approach maintains training accuracy by preserving the critical directional information while significantly reducing the computational and memory burden on resource-constrained edge devices.

Inventive Principle:
Principle #2Taking out (Extraction)

Data Source

PatentEP4664324A1Method and system for training a distributed model
Publication Date: 2025.12.17 TECHNISCHE UNIVERSITAT MUNCHEN
  • EP4664324A1 patent drawingFigure 1~2
  • EP4664324A1 patent drawingFigure 3~4
  • EP4664324A1 patent drawingFigure 5a~5b

AI summary

The present disclosure provides computer-implemented method for training a model using a central node and a plurality of client nodes, comprising, at one or more of the plurality of client nodes: receiving a set of model parameters and an initial seed from the central node; determining an update direction based on the initial seed; computing scalar reflecting a sign and a size of an approximation of a projection of a gradient of a loss function on the update direction, using the set of model parameters; sending the scalar of the approximated gradient projection to the central node; receiving a model update scalar from the central node; and updating the set of model parameters in the update direction based on the model update scalar.