Network Certifiers for Inference Verification - Technical Overview

Introducing a physical prototype for AI verification using zero-knowledge proofs and a FPGA

November 2026

Physical prototype

Inference verification is the process of ensuring that datacenters are running inference workloads as claimed. Verifiable inference can enable AI deployment: allowing consumers to confirm which model is being deployed in order to increase their confidence in the models. Verification is also an important governance tool. It can give governments assurance that the deployed model is the same as the one they evaluated and let them confirm that datacenters are not running unauthorized workloads, such as the training of dangerous AI models.

SASH and Future of Life Institute (FLI) have jointly prototyped a device that fingerprints network traffic entering and leaving a datacenter: a Network Certifier. Using Zero-Knowledge Proofs (ZKPs), we built a system to analyze these fingerprints, allowing us to verify the correctness of inference workloads while preserving full confidentiality of sensitive data, including model weights, inputs and output.

In this post, we summarize our approach and link to the technical reports on the Network Certifier protocol and ZKP-based verification.

Component diagram

Concept Overview

The goal of our system is to verify that all data leaving a GPU cluster are the result of correct inference on the declared model. Inference can be incorrect if the model is substituted or tampered with during deployment, user inputs are modified, workloads include malware aimed at exfiltrating model weights, or if the cluster is being used for unauthorized training instead of inference.

Our system consists of two components aimed at detecting these violations: fingerprint collection and fingerprint analysis. A fingerprint is a cryptographic hash corresponding to specific input and output data, and the committed model weights.

Fingerprint Collection

The collection of fingerprints is performed by an auditor device known as a Network Certifier. This device is retrofitted into each GPU cluster so that all network inputs and outputs must pass through it. When traffic passes through, the Network Certifier collects cryptographic fingerprints of all traffic that passes through. Since traffic is encrypted by the datacenter operator before leaving the datacenter, neither the auditor nor their devices can ever see the raw data. This ensures the confidentiality of both model weights and user data.

Fingerprint Analysis

The auditor now has the fingerprints of the inputs and outputs. In addition, as part of the model approval process, the datacenter operator also fingerprints their model weights and declares them to the auditor.

By analyzing these input, output and model fingerprints, our system is able to determine if the datacenter is running the correct workload. This analysis is done using ZKPs.

Fingerprint analysis

If the datacenter runs the declared model correctly, the input, output and model fingerprints will be consistent. If the workload was performed incorrectly – for example if the user prompt was modified by the datacenter operator – the output fingerprints would change and become inconsistent with the input and declared model fingerprints.

By using ZKPs, our system allows the datacenter operator to prove the consistency of the fingerprints to an auditor without having to reveal any other information.

Implementation

The Network Certifier is built on the MPF-300TS PolarFire FPGA using the protocol described here. The fingerprint analysis is built on top of the Ligero ZKP system and relies on the techniques described in VerInf and Verifying AI Compute by Bounding Unexplained Information Exfiltration.

Proving all inference requests using ZKPs would be impractically slow. To reduce the performance costs of verifying inference, auditors can prove a randomly selected sample of inputs. Additionally, we make use of sub-sampling – proving a subset of the layers in a network instead of the whole thing – to further reduce costs. More details about the system can be found below.

Technical Reports:

Next Steps

In the coming months we plan to build on our existing research in zero-knowledge proof — in particular exploring sub-sampling methods and integrating provers with LLM inference engines — and resource exhaustion verification approaches, as well as our policymaker engagement.

Interested in getting involved? Apply to one of our open roles Submit improvements to the codebases linked above