Zero-Knowledge LLM Identification From Physical Telemetry

A groundbreaking research paper published today on Zenodo introduces a novel method for identifying large language models through physical telemetry analysis while preserving privacy through zero-knowledge cryptographic techniques. The work addresses a growing need in AI verification: confirming which model is running on hardware without accessing its internal parameters or architecture.

Physical Telemetry as Model Fingerprints

The research demonstrates that different LLM architectures produce distinctive patterns in physical signals emitted during inference. Power consumption profiles, electromagnetic radiation, and thermal signatures vary predictably based on model size, architecture type, and optimization strategies. A 70-billion parameter model exhibits markedly different power draw patterns compared to a 7-billion parameter variant, even when processing identical prompts.

The researchers collected telemetry data across multiple hardware platforms, including NVIDIA A100 and H100 GPUs, finding consistent fingerprinting accuracy above 94% for major model families. These physical side-channels reveal information about computation patterns, memory access sequences, and attention mechanism implementations that remain consistent across inference sessions.

Zero-Knowledge Verification Framework

The core innovation lies in combining physical telemetry with zero-knowledge proof systems. Traditional model identification methods require direct access to weights, activation patterns, or API responses. This approach enables third parties to verify model identity without the model operator revealing proprietary information about architecture, training data, or fine-tuning procedures.

The zero-knowledge protocol allows a prover (the entity running the LLM) to demonstrate they are executing a specific model version while keeping all implementation details confidential. The verifier analyzes only aggregate telemetry statistics rather than raw inference data, preserving both user privacy and model intellectual property.

Real-Time Implementation Challenges

The system operates with latency under 200 milliseconds for model identification, making it suitable for compliance monitoring and API verification scenarios. However, the researchers note several practical constraints:

• Physical access to telemetry sensors is required, limiting applicability to cloud environments without hardware instrumentation
• Highly optimized or quantized models may produce overlapping signatures with different architectures
• Adversarial operators could potentially spoof telemetry patterns through deliberate power consumption manipulation
• Cross-platform calibration requires extensive baseline measurements for each hardware configuration

The methodology shows particular promise for regulatory scenarios where independent auditors need to verify that deployed systems match certified model versions without accessing proprietary code or training pipelines.

Release Date and Availability

The research paper was officially released on October 31, 2026, through the Zenodo open science repository. The authors have made their dataset of telemetry measurements publicly available for reproducibility, covering 18 popular open-source and commercial LLM architectures across four hardware platforms.

The accompanying codebase includes reference implementations of both the telemetry collection framework and the zero-knowledge verification protocol, licensed under Apache 2.0 for commercial and academic use.

What This Means

This research opens new possibilities for AI system auditing and compliance verification in environments where traditional white-box inspection is impractical or prohibited. As LLM deployment scales into regulated industries like healthcare and finance, independent verification methods that respect intellectual property become increasingly critical. The fusion of physical side-channel analysis with cryptographic privacy guarantees represents a pragmatic path toward trustworthy AI infrastructure, though widespread adoption will require standardized telemetry collection interfaces across cloud providers and edge deployment platforms.