R01: Dash: Accelerating Distributed Private Convolutional Neural Network Inference with Arithmetic Garbled Circuits

Status: reference primitive only
Year: 2024
Authors: Jonas Sander, Sebastian Berndt, Ida Bruhns, Thomas Eisenbarth
External source: Dash: Accelerating Distributed Private Convolutional Neural Network Inference with Arithmetic Garbled Circuits

Source contribution

Arithmetic garbling and LabelTensors provide the foundation for offline preparation and local encoded public-weight evaluation.

PLLM perspective

Implement affine-label arithmetic, constant multiplication with the required algebraic preconditions, projections and tensorized evaluation. Preserve the original circuit functionality before adapting to transformers.

Prior single-server, model-aware and tenx archives contain adaptations, not a reproduced Dash release.

PLLM has related clean-room reference primitives only. They are not a paper reproduction, reviewed protected runtime, or complete-model result.

The source's boundary remains binding: Do not inherit malicious security by importing kernels: the paper discusses additional integrity mechanisms and trusted-hardware or cryptographic configurations.

PLLM modules and documentation

Target module identifiers: garble.affine_label, kernel.label_tensor, gate.projection. These identifiers describe scoped work unless this page explicitly links an implementation source. Dependencies: R03, R04.

• Research backlog
• Reimplementation backlog
• Clean-room workflow
• Arithmetic garbling documentation
• Reference garbling source

Backlog

This source is assigned to reimplementation priority 3, Arithmetic garbling foundation. Its source-registry order is 1. Follow current gates and unfinished work in the research backlog.