Research
Privacy-Preserving AI Verification via Minimal Information Disclosure
Verifying claims about AI execution — that a data center trained a model, stayed within a compute limit, used an approved accelerator, or ran a particular model type — requires evidence such as power…

- arXiv
- 2608.02774
- Published
- 2026-08-03
- Authors
- Sleem Abdelghafar, Gabriel Kulp
AI summary
Overview
Research area: Cryptography and Security (AI verification, information-theoretic privacy, zero-knowledge proofs, physical side channels).
Technical level: Intermediate. The paper combines mutual-information arguments with concrete measurement experiments and zk-SNARK tooling; readers comfortable with basic information theory and applied cryptography will follow it, and the empirical sections are largely readable without either.
Scope (one sentence): The paper proposes minimal information disclosure (MID), a framework that chooses which verifier-facing evidence mechanism to deploy so that an AI execution can be verified while leaking as little as possible about a declared protected property beyond the authorized answer.
What This Paper Is About
Verifying claims about AI execution — that a data center trained a model, stayed within a compute limit, used an approved accelerator, or ran a particular model type — requires evidence such as power traces, telemetry, or electromagnetic measurements. That same evidence can also expose workload details, model architecture, or hardware configuration that the verifier has no right to learn. The paper's goal is to make the information content of the evidence itself the design object: pick, from a set of deployable measurement-and-release mechanisms, the one that preserves enough information to confirm the authorized claim while minimizing "collateral" disclosure about a protected property, measured as conditional mutual information I(S; E | Y) once the authorized result Y is known.
Key Contributions
-
An information-theoretic formulation of privacy-preserving AI verification. The paper writes the problem as choosing a mechanism m = (c, α, θ) — evidence channel c, collection policy α, optional release transformation θ — to minimize I(S; E_m | Y) subject to I(Y; E_m) ≥ τ_Y, with an extension to multiple sensitive targets that minimizes the maximum leakage across them. The formulation is designed to generalize across verification goals, protected properties, evidence sources, and deployment constraints.
-
Two deployment paths built from concrete mechanism families. Collection-time mechanisms (channel restriction, acquisition minimization) let a sensor produce a low-information measurement directly, so no richer trace is ever created. Post-collection mechanisms (hard decisions, randomized quantization, and a privacy-aware linear projection that penalizes directions separating S within fixed Y) keep the rich measurement inside a protected boundary and release only a transformed value, certified by a zero-knowledge proof of correct computation. The linear-projection arithmetic is demonstrated in Circom with a Groth16 zk-SNARK verified over BN254 (called bn128 by snarkjs).
-
Empirical demonstration across six verification tasks and four physical-measurement datasets. The tasks span execution type, hardware identity, compute scale, and model identity. Three selected mechanisms achieve perfect held-out verification with zero measured collateral leakage; the remaining tasks produce explicit privacy–utility frontiers. Selected releases are also evaluated against the ModelSpy and DeepTheft attack methods.
-
Evidence that output minimization alone is insufficient. Using Auditor-in-a-Box's released evaluation, the authors show that a predictor receiving only that system's one-bit valid/invalid verdict — not the request itself — identifies which of the authors' two request sets a private request came from with 72.4% balanced accuracy, compared with 50% without the decision.
Main Findings
-
Training versus inference (NLR dataset): Raw facility-power features verified training-vs-inference perfectly but supported sensitive-workload inference at mean balanced accuracy 0.9182, with I(Y; E) = 1.0000 bits and measured conditional leakage 0.6375 bits. The selected MID report — six dimensionless shape statistics (coefficient of variation, peak-to-mean ratio, autocorrelation at lags 1, 5, and 10, and spectral entropy), excluding absolute power, duration, energy, sample count, and sampling interval, released through a randomized quantizer with q = 0.05 and σ = 0.15 — retained authorized balanced accuracy 0.9966 and I(Y; Z) = 0.9723 bits while dropping mean sensitive-workload accuracy to 0.5010 and measured conditional leakage to 0.000239 bits.
-
Per-class sensitive accuracy collapses to chance: In that same task, sensitive balanced accuracy fell from 0.9750 to 0.5000 within training and from 0.8613 to 0.5020 within inference.
-
The one-bit verdict still leaks: Auditor-in-a-Box reports a valid/invalid verdict per researcher specification intended as a one-bit worst-case disclosure bound; the authors' predictor still separates the two accepted-request sets at 72.4% balanced accuracy versus a 50% baseline, on the authors' own evaluation distribution.
-
RL versus non-RL (Elsayed dataset): A rich baseline of 144 GPU/CPU power, utilization, memory, temperature, and voltage summaries recovers both the RL/non-RL target and the four-way exact non-RL workload. MID's selected channel uses only three fields — GPU core load, activity through NVIDIA's CUDA computing interface, and mean Direct3D utilization — and releases a single thresholded bit of the session-level RL probability. The numerical results for this task appear after the truncation point of the supplied text and are not reported here.
-
Hardware identity, compute scale, and model identity tasks: The paper lists these tasks (H100 vs B200 from accelerator telemetry; low vs high execution tier and node/GPU counts from processor-package, memory, and facility power; CNN vs Transformer from GPU electromagnetic traces). Their results lie in Sections 4.3–4.5, outside the supplied truncated content, and are not reported here.
-
Attack-based checks: The paper states that selected releases are evaluated against ModelSpy and DeepTheft, but the outcomes are not present in the supplied text.
-
Theoretical guarantee: Because any downstream inference rule A computed from (E_m, Y) satisfies the conditional Markov relation S → E_m → A given Y, the conditional data-processing inequality gives I(S; A | Y) ≤ I(S; E_m | Y), so the reported leakage bounds any later attack on the specified S. Fano's inequality then converts a leakage value into a lower bound on an adversary's error probability.
Methodology in Plain English
The workflow starts with a policy decision: which claim may the verifier check (Y), which property must stay hidden (S), and how informative must the evidence be (threshold τ_Y). During development the authors have access to a rich recording of each execution, from which they can simulate many candidate mechanisms — measuring a different signal, sampling less often, or post-processing the reading into something coarser or noisier.
Each candidate is scored on two quantities estimated from data: how much it tells the verifier about Y (mutual information, or balanced accuracy in practice), and how much it tells the verifier about S after Y is already known (conditional mutual information). Candidates that fail the utility threshold are discarded; among the rest, the one with the least leakage wins. To avoid flattering themselves, the authors split the data twice — inner folds to compare candidates, and an outer set of held-out physical runs, grouped so windows from the same run never appear on both sides — to check the whole selection procedure. For continuous traces they use a regularized Gaussian separation score (a Mahalanobis distance with a stabilization term) as a finite-sample stand-in, plus held-out prediction and attack tests to probe structure the Gaussian model may miss.
Deployment has two routes. If a cheaper sensor or collection rule already meets the utility threshold, the sensor simply emits that low-information reading, authenticated and bound to the execution. If a post-collection transformation is better, the raw measurement stays inside a sensor, secure meter, or TEE, the source commits to it and signs the commitment, and a zk-SNARK proves that the released value was computed correctly from the committed private input — the verifier sees the execution identifier, commitment, released value, mechanism parameters, source signature, and proof, but not the raw trace or the randomness.
Why This Matters
Verification regimes for AI — compute governance, accelerator end-use controls, inference-only policies, model attestation — all hinge on evidence crossing a trust boundary. The paper's point is that deciding how much output a verifier receives (a single bit, a yes/no verdict) is not the same as deciding how much it learns. MID provides a way to compare candidate evidence designs quantitatively before deployment, and to state the privacy–utility tradeoff explicitly when no zero-leakage design exists.
Real-world applications implied by the paper's tasks:
- Compute governance and international verification: demonstrating that a training run stayed within an agreed compute class, or that only inference occurred, using facility power rather than detailed workload records.
- Export control and chip licensing: attesting that an approved accelerator (e.g., H100 vs B200) was the one in use without revealing the customer's model or workload.
- Model and property attestation: confirming that a claimed model family (CNN vs Transformer) executed, without disclosing architecture specifics to the auditor.
- Compliance reporting under disclosure constraints: converting rich telemetry into narrow, ZK-certified reports so that a regulator receives a compliance determination rather than proprietary traces.
Industry relevance: the framework gives hardware vendors, cloud operators, and auditors a design methodology for "verifier-facing" interfaces, and it slots in alongside authentication, hardware attestation, and existing ZKP systems rather than replacing them — MID decides what the report should say, while those tools establish where it came from and how it was computed.
Future Directions
-
Scaling mechanism search beyond three controls. The authors explicitly note MID is not limited to channel, collection policy, and release transformation, and can accommodate other deployable mechanisms by extending the candidate set M. How large a mechanism library remains searchable with finite calibration data is open.
-
Extending to richer evidence and more sensitive targets. The experiments use physical measurements; the formulation also covers logs, telemetry records, and stored models. Evaluating the multi-target min-max form, and the case where the combination of S₁, …, S_J is itself sensitive, is left to future work.
-
Deployment hardening of the proof path. Groth16 requires a circuit-specific structured reference string, and randomized releases require verifiable sampling (a committed hidden seed plus verifier challenge). The paper describes these requirements but the supplied content does not report a full end-to-end deployment cost or setup ceremony.
-
Adversarial and adaptive verifiers. MID bounds information under the declared protected target S; how well the guarantee holds when the verifier's real interest differs from, or is broader than, the declared S — and how the Auditor-in-a-Box-style residual leakage scales with verdict alphabets larger than one bit — is a natural next question.
Target Audience
Researchers and practitioners working on AI governance and verification, compute governance, hardware attestation, and side-channel privacy; cryptographers interested in how ZK proofs can be paired with information-theoretic release design; and policy or standards groups who need to reason about what a compliance report actually discloses. Readers focused purely on ML systems performance will find the empirical sections accessible, while the mechanism-design and proof sections are aimed at security engineers.
Authors’ abstract
AI verification crosses a trust boundary: a verifier must learn enough to establish an authorized claim, yet the same evidence can reveal sensitive details about the model, workload, or hardware. We introduce minimal information disclosure (MID), which designs and quantifies the information content of verifier-facing evidence itself. MID measures collateral leakage with conditional mutual information: what the release reveals about the protected property after the authorized…
The full abstract on arXiv