P093 Workload-type classification from power and telemetry
How accurately can classifiers separate training from inference from non-AI work using telemetry or power, and does that accuracy survive an adaptive evader and hardware drift?
Accuracy, evasion robustness and generalisation of classifiers that tell training from inference from non-AI work using on-board telemetry or off-chip/rack/facility power.
Terms: β (beta): the largest hidden computation a power trace cannot rule out, as a fraction of the declared machine capacity; β (beta) is its worst case over declared utilisation (how busy the declared work keeps the chip).
Status
Started External power draw of a single NVIDIA H200 NVL sampled at 10 MHz, 5 s traces, 17 LLM families and 25 non-AI workloads, authors' own classifier; four evasion strategies tested by the authors, who say this is not a comprehensive adversarial evaluation: three caught at or above 99%, diluted LoRA at 48-88% without the rescue rule, below this survey's detection goal of 90%; cross-device and cross-generation generalisation 'remains to be established'; single node. Corroborating own-method report: Rahman & Tajdari (2026) classify training against non-training from zero-overhead GPU telemetry at 98.2% across 9 GPU types, single node up to 8 GPUs, with dilution and idle-padding evasion tested. Harack et al. (2025) 'not yet possible to reliably know' predates both.
"we separate training from inference and from non-AI computation with an accuracy of 97% and a macro-averaged F1 score of 0.955, evaluated on model families unseen during training" … "The fourth, diluted low-rank adaptation (LoRA), is detected 48–88% of the time with a hardened classifier"
Posed by
- Scher & Thiergart (2024), Mechanisms to Verify International Agreements About AI Development (MIRI):
Many of these methods appear easily spoofable given the present threat model, e.g., by making training 20% slower in order to look like inference, within the budget of relevant actors.
(Known Compute › Workload Classification, p. 26, L1198–1200)
Also raised in: Reuel, Bucknall et al. (2025), Open Problems in Technical AI Governance, compute sections; Harack et al. (2025), Verification for International AI Governance (Oxford AIGI); Baker, Kulp, Marks, Brundage & Heim (2025), Verifying International Agreements on AI: Six Layers of Verification for Rules on Large-Scale AI Development and Deployment (RAND WR-A4077-1); Petrie & Aarne (2025), Technical Options for Flexible Hardware-Enabled Guarantees (flexHEG Part 2); Shavit (2023), What Does It Take to Catch a Chinchilla? Verifying Rules on Large-Scale Neural Network Training via Compute Monitoring; Scher (2026), Verifying Restrictions on Frontier AI Research (MIRI); Canary Institute research; Abdelghafar & Kulp (2026), Privacy-Preserving AI Verification via Minimal Information Disclosure; Ansari (2026), Hardware-Level Governance of AI Compute: A Feasibility Taxonomy for Regulatory Compliance and Treaty Verification; Gargiulo & Kulp (2026), Workload Identification with Physical Side Channels for AI Governance; Hausenloy & Li (2026), The Compute Verification Post; Koopmanschap & Barten (2026), How to Catch a GPU: A Taxonomy of Verification and Enforcement Mechanisms for International AI Agreements; Kulp et al. (2024), Hardware-Enabled Governance Mechanisms (RAND WR-A3056-1); Moon, Vedula, Geneson & Bar-on (2025), Strategies and Detection Gaps in a Game-Theoretic Model of Compute Governance (RAND RRA3686-1); Rahman & Tajdari (2026), Detecting Hidden ML Training With Zero-Overhead Telemetry (MIRI); Wasil, Reed, Miller & Barnett (2024), Verification Methods for International AI Agreements; Kimpson, Baker & Graham (2026), Can Power Draw Constrain Covert Compute? Limits of Analogue Verification for AI Governance, arXiv 2610.07476; AI Security Forum, Requests for Discussion; Proofworks, Living Verification Documents.
Sub-questions
| ID | Sub-question | What the sources report |
|---|---|---|
| P093.1 | Does any governance proposal give a numeric performance spec (e.g. max undetected FLOP fraction) for the off-chip analog/physical layer it relies on? | open |
| P093.2 | With cooperative access (prover not hiding) to an 8-GPU node, to what accuracy can rack-level power alone classify training vs inference vs non-ML? | partly answered |
| P093.3 | Does any off-chip (non-NVML) analog workload classifier keep a stated accuracy against an adaptive evader in a published red-team evaluation? | open |
| P093.4 | Can Maya-style active power shaping make a training workload's rack-power trace pass an inference classifier at a stated throughput cost? | open |
| P093.5 | Does adding an EM (or acoustic/thermal) channel to the power meter lower the detection threshold β* by a stated factor? | open |
| P093.6 | Using only a high-pass band above a stated cutoff (amplitude-normalised), what training-vs-inference accuracy does the power classifier reach? | partly answered |
| P093.7 | By what factor does accelerator-level masking cut workload-classification accuracy from off-chip power? | partly answered |
| P093.8 | Does a telemetry-based training classifier on AMD ROCm SMI or Huawei npu-smi reach a stated detection rate? | partly answered |
| P093.9 | Can an inspector with no chip-level activity logs determine from facility records whether a training run above a stated compute threshold occurred? | implied by a source, not stated as open |
| P093.10 | Has any architecture or hyperparameter been recovered from EM/power of a data-centre GPU (A100/H100) in a populated rack? | open |
| P093.11 | What workload-type accuracy (incl. training vs inference) do rack/PDU/facility power traces give on held-out AI jobs? | partly answered |
| P093.12 | What is the false-positive rate (AUC) of a power-profile training detector on non-AI HPC traces (climate, MD, finance)? | open |
| P093.13 | Can load-shape features separate AI training facilities from crypto-mining or other continuous industrial loads at the utility meter? | open |
| P093.14 | Can a red team reproduce a declared workload's power/I-O signature using a stated fraction of the declared chip-hours? | open |
| P093.15 | On a held-out workload mix, what training / inference / non-AI classification accuracy does hardware telemetry alone (power, utilisation, memory bandwidth, counters) reach without an adversary? | partly answered |
| P093.16 | Can chip-level code, memory snapshots or compiler signatures separate neural-network training from other HPC jobs (e.g. simulation) with a stated TPR at fixed FPR? | open |
| P093.17 | Against an adaptive evader who shapes power/utilisation to mimic inference or a named non-AI job, what detection rate does the best telemetry classifier retain? | partly answered |
| P093.18 | What throughput penalty must an evader pay to push a classifier's detection below a stated level, and does a white-box evader always win? | partly answered |
| P093.19 | Does any classifier reach a stated detection rate (a target to be agreed) for diluted LoRA or other parameter-efficient / reshaped full fine-tuning disguised as inference? | partly answered |
| P093.20 | Can custom CUDA kernels, matrix decomposition or computation reordering drop hardened-classifier detection below a stated floor (a target to be agreed)? | open |
| P093.21 | Does classifier accuracy hold when the evader controls or times around the telemetry read path (NVML values, unsampled windows)? | partly answered |
| P093.22 | What minimum sampling rate and observation window keep training-vs-inference detection at a stated high level (a target to be agreed) under evasion? | a source suggests an answer (not checked) |
| P093.23 | Does a classifier trained on one GPU SKU/software generation keep accuracy (drop within a stated bound to be agreed) on other devices, architectures, vendors (incl. mixed NVIDIA/Huawei clusters) and over weeks of drift? | open |
| P093.24 | Does classifier accuracy and evasion detection hold on multi-node InfiniBand (node and GPU counts a target to be agreed) frontier-scale runs and larger models? | open |
| P093.25 | What false-alarm rate does the classifier (incl. rescue rules) produce on production multi-tenant serving, ETL and other uncovered workloads? | partly answered |
| P093.26 | Can RL post-training, fine-tuning and continual learning be separated from batch inference and from full pre-training at a stated AUC (a target to be agreed)? | open |
| P093.27 | Are per-workload signatures specific enough to tell the same model at different scales or to match a workload against its expected fine-grained physical signature? | open |
| P093.28 | Ranked by classification accuracy on a labelled dataset, which channel (power, EM, network, timing, energy-per-token) is most informative, and can network features alone reach a stated high accuracy (a target to be agreed)? | open |
| P093.29 | Can EM emanations identify which model or workload runs on a datacenter GPU despite software obfuscation? | a source suggests an answer (not checked) |
| P093.30 | Can provider-side utilisation telemetry or customer metadata flag or size (to a stated FLOP precision to be agreed) above-threshold runs without seeing code or data? | open |
| P093.31 | Does a public labelled dataset of compliant vs non-compliant accelerator workload traces exist, and of what size? | open |
Sub-questions not yet stated precisely enough to check
| ID | Sub-question |
|---|---|
| P093.32 | What do candidate sensors actually measure (bandwidth, resolution, placement limits)? |
| P093.33 | Can a labelled signal dataset remain useful across hardware generations? |
| P093.34 | Are verifier-derived signatures from similar hardware robust to backdoored hardware? |
Possibly relevant work found by automated search
Found by AI agents searching for this problem's sub-questions, mostly by keyword and citation in the OpenAlex scholarly database and on the web; a second AI pass reviewed each one, and we spot-checked some by hand. They seemed reasonably related, as far as one can tell from outside each paper's field. Some come from neighbouring fields, such as arms control and financial audit, and bear on the problem by analogy. Not reviewed against the status above; a paper listed here does not change that status. 928 of 1052 quotes checked word for word against the paper's text.
For P093.1: Does any governance proposal give a numeric performance spec (e.g. max undetected FLOP fraction) for the off-chip analog/physical layer it relies on?
For P093.2: With cooperative access (prover not hiding) to an 8-GPU node, to what accuracy can rack-level power alone classify training vs inference vs non-ML?
For P093.4: Can Maya-style active power shaping make a training workload's rack-power trace pass an inference classifier at a stated throughput cost?
For P093.5: Does adding an EM (or acoustic/thermal) channel to the power meter lower the detection threshold β* by a stated factor?
For P093.6: Using only a high-pass band above a stated cutoff (amplitude-normalised), what training-vs-inference accuracy does the power classifier reach?
For P093.8: Does a telemetry-based training classifier on AMD ROCm SMI or Huawei npu-smi reach a stated detection rate?
For P093.10: Has any architecture or hyperparameter been recovered from EM/power of a data-centre GPU (A100/H100) in a populated rack?
For P093.11: What workload-type accuracy (incl. training vs inference) do rack/PDU/facility power traces give on held-out AI jobs?
For P093.12: What is the false-positive rate (AUC) of a power-profile training detector on non-AI HPC traces (climate, MD, finance)?
For P093.15: On a held-out workload mix, what training / inference / non-AI classification accuracy does hardware telemetry alone (power, utilisation, memory bandwidth, counters) reach without an adversary?
For P093.16: Can chip-level code, memory snapshots or compiler signatures separate neural-network training from other HPC jobs (e.g. simulation) with a stated TPR at fixed FPR?
For P093.17: Against an adaptive evader who shapes power/utilisation to mimic inference or a named non-AI job, what detection rate does the best telemetry classifier retain?
For P093.18: What throughput penalty must an evader pay to push a classifier's detection below a stated level, and does a white-box evader always win?
For P093.21: Does classifier accuracy hold when the evader controls or times around the telemetry read path (NVML values, unsampled windows)?
For P093.23: Does a classifier trained on one GPU SKU/software generation keep accuracy (drop within a stated bound to be agreed) on other devices, architectures, vendors (incl. mixed NVIDIA/Huawei clusters) and over weeks of drift?
For P093.28: Ranked by classification accuracy on a labelled dataset, which channel (power, EM, network, timing, energy-per-token) is most informative, and can network features alone reach a stated high accuracy (a target to be agreed)?
For P093.30: Can provider-side utilisation telemetry or customer metadata flag or size (to a stated FLOP precision to be agreed) above-threshold runs without seeing code or data?
Data
- problems.csv — one row per problem
- problem_status.csv — status, scope and evidence per problem
- problem_sources.csv — the posed-by quote and other documents per problem
- leaves.csv — all sub-questions, including those without a checkable form
- leaf_sources.csv — every source row behind each sub-question