Black-Box Model Auditing

Ask about this topic

Overview

Black-box model auditing studies what can be verified about an AI system when its weights, training data, or internal processes are inaccessible. Auditors use statistical tests and membership inference, active or sequential query selection, counterfactual and causal probes, behavioral fingerprints, and sometimes cryptographic verification to examine privacy, fairness, safety, robustness, explanations, and model identity. These methods are applied to settings from hosted language-model APIs and recommendation systems to scientific predictors, often using controlled benchmarks, matched interventions, and repeated trials. A central tension is that limited or manipulable outputs can conceal rare failures and make latent properties hard to establish, so audit conclusions depend on query access, measurement reliability, and how closely tests reflect deployment.

Latest papers

CardsList

96 papers11 papers in the last four weeks, down 35% on the four weeks before. 0.1% of all new papers.

  1. Fundamental Limits of Black-Box Safety Evaluation: Information-Theoretic and Computational Barriers from Latent Context Conditioning

    Feb 19, 2026Vishal SrivastavaAI Safety EvaluationBlack-Box Model Auditing

    Whiting School of Engineering, John Hopkins University, Baltimore, MD 21218

  2. Audit Me If You Can: Query-Efficient Active Fairness Auditing of Black-Box LLMs

    Jan 6, 2026David Hartmann, Lena Pohlmann, Lelia Hanslik +3LLM AuditingAlgorithmic Fairness

    Weizenbaum Institut Berlin · Technische Universität Berlin · FIZ Karlsruhe +2

  3. Data Provenance Auditing of Fine-Tuned Large Language Models with a Text-Preserving Technique

    Oct 7, 2025Yanming Li, Cédric Eichler, Nicolas Anciaux +3Data ProvenanceDataset Watermarking

    PETSCRAFT · Petscraft, Inria, Palaiseau, France · INSA CVL, Bourges, France +6

  4. CAFÉ: Causal Black-Box Testing of Machine Unlearning

    Sep 20, 2025Anna Mazhar, Sainyam GalhotraMachine Unlearning EvaluationBlack-Box Model Auditing

    Cornell University, USA

  5. Optimizing Canaries for Privacy Auditing with Metagradient Descent

    Jul 21, 2025Matteo Boglioni, Terrance Liu, Andrew Ilyas +1Privacy AuditingDifferentially Private Stochastic Gradient Descent

    Carnegie Mellon University · Stanford Statistics

  6. CBW: Towards Dataset Ownership Verification for Speaker Verification via Clustering-based Backdoor Watermarking

    Mar 2, 2025Yiming Li, Kaiying Yan, Jiawen Diao +4Speaker VerificationDataset Watermarking

    College of Computing and Data Science, Nanyang Technological University, Singapore, 639798, Singapore · School of Artificial Intelligence, Shanghai Jiao Tong University, Shanghai, 200030, China · Beijing Electronic Science and Technology Institute, Beijing, 100070, China +4