Fine-tuning LLMs on privacy-sensitive data introduces privacy risk, and synthetic data audits can quantify this risk.
problem Fine-tuning LLMs on privacy-sensitive data introduces privacy risk.
method Generate synthetic canaries via high-temperature sampling from LLMs.
result Synthetic canaries are high-influence outliers that ensure strong audits.
Improved canary crafting for one-run privacy auditing reduces leakage estimates.
problem Detecting canaries in one-run privacy auditing to estimate leakage effectively.
method Optimizes canaries for detectability and diversity, using a greedy initialization and bilevel optimization.
result Achieves stronger leakage estimates at lower computational cost.
New divergences help audit DP in high dimensions.
problem Challenges in auditing DP in high-dimensional data.
method Propose kernel Rényi divergence and its regularized version for auditing.
result Regularized kernel Rényi divergence can be estimated from samples in high dimensions.
Big data transforms accounting and auditing, enhancing insights but posing challenges.
problem Challenges in data privacy and security with increased data sources.
method Utilizing AI and machine learning for efficient data analysis and anomaly detection.
result Enhanced analytics tools and continuous learning are key to overcoming challenges.
Privacy concerns have led to the development of privacy-preserving approaches for learning models from sensitive data. Yet, in practice, even models learned with privacy guarantees can inadvertently memorize unique training examples or leak sensitive features. To identify such privacy violations, existing model auditin…
New method audits DP guarantees without noise or subsampling info.
problem Auditing DP guarantees of ML models without prior info.
method Histogram-based density estimation for lower bounds.
result Natural generalization of membership inference auditing.
New method protects whistleblowers from retaliation by ensuring their reports remain private.
problem Whistleblowers face retaliation, and current protections are insufficient.
method Formalizes protection against strong-adversary threat model as per-report (0,δ)-differential privacy, and provides a generic mechanism to reduce private auditing to private continual counting. result Demonstrates a reduction in selection error and improved utility over randomized response.
Synth-MIA assesses privacy leakage in synthetic tabular data models.
problem Challenges in evaluating privacy leakage in synthetic tabular data.
method Unified threat framework deploying multiple attacks.
result Higher synthetic data quality correlates with greater privacy leakage.
BlockFLow ensures privacy and accountability in federated learning.
problem Malicious agents can weaken federated learning models.
method Differential privacy, auditing mechanism, Ethereum smart contracts.
result Audit scores reflect the quality of honest agents' datasets.
New attacks improve privacy audits by analyzing model updates.
problem Improving privacy audits of AI models through sequence analysis.
method Developed SeMI attacks to identify target insertions in model sequences.
result SeMI attacks achieve higher power and tighter privacy audits.
New auditors assess f-DP privacy with adaptive sampling, avoiding large sample sizes.
problem Empirical auditing of f-DP privacy with adaptive sampling. method Shift focus to f-DP, develop adaptive auditors for whitebox and blackbox settings. result Adaptive auditors detect f-DP violations across the privacy spectrum with statistical guarantees. Framework audits synthetic datasets for trustworthiness across various use cases.
problem Assessing the trustworthiness of synthetic datasets and models.
method Holistic auditing framework focusing on bias, fidelity, utility, robustness, and privacy.
result Introduces a trustworthiness index and model selection process for controllable trade-offs.
Gen-LRA attacks synthetic data leakage without model knowledge.
problem Auditing synthetic data privacy leakage.
method Generative Likelihood Ratio Attack (Gen-LRA).
result Gen-LRA outperforms other attacks across metrics.
A privacy-preserving synthetic data generation framework that distinguishes between true and phantom data disclosures.
problem Detecting and explaining data disclosures in synthetic datasets.
method Customizable empirical auditing framework with statistical hypothesis testing.
result Demonstrated tighter privacy leakage bounds than prior methods.
A new metric GNQ audits LLMs for privacy risks during training.
problem Auditing LLMs for privacy risks during training is computationally hard.
method Gradient Uniqueness (GNQ) metric derived from gradient descent, BS-Ghost GNQ for efficiency.
result GNQ successfully predicts sequence extractability and reveals risk heterogeneity.
Privacy-preserving machine learning methods add randomness, leading to varying predictions.
problem Privacy-preserving machine learning methods add randomness, leading to varying predictions.
method The study analyzes three DP-ensuring algorithms: output perturbation, objective perturbation, and DP-SGD.
result The degree of predictive multiplicity rises as the level of privacy increases, and is unevenly distributed across individuals and demographic groups.
Paper shows incorrectness of approximate unlearning definitions and challenges exact unlearning verification.
problem Incorrectness of approximate unlearning definitions and challenges in verifying exact unlearning.
method Analysis of machine unlearning approaches, including exact and approximate methods.
result Unlearning is only well-defined at the algorithmic level, and auditable claims are limited.
This paper operationalizes the Exponential Mechanism using Normalizing Flows for private optimization.
problem Improving privacy in machine learning while maintaining accuracy and efficiency.
method Using Normalizing Flows to approximate sampling from the Exponential Mechanism for private optimization.
result ExpM+NF provides more privacy than non-private SGD but not as much as DPSGD.
Accurately learning from user data while providing quantifiable privacy guarantees provides an opportunity to build better ML models while maintaining user trust. This paper presents a formal approach to carrying out privacy preserving text perturbation using the notion of dx-privacy designed to achieve geo-indistingui…
Study proposes BFEL framework for privacy-preserving FL in personalized healthcare.
problem Privacy and security concerns in traditional cloud-centric ML, especially in wearable devices.
method Develops a blockchain-enhanced federated edge learning (BFEL) framework based on FedCurv, incorporating fisher information matrix and public key encryption.
result Significant reduction in communication cost and high efficiency for federated training on non-iid and heterogeneous data.
New MCMC method estimates differential privacy from multiple MIAs without worst-case assumptions.
problem Bayesian estimation of differential privacy from membership inference attacks.
method Bayesian estimation via MCMC algorithm (MCMC-DP-Est).
result More cautious privacy analysis with joint estimation of MIA strengths and privacy parameter.
Log-Loss scores expose membership privacy breaches.
problem Privacy leakage from statistical aggregates like Log-Loss scores.
method Proved that Log-Loss scores enable full accuracy membership inference in a single query.
result Complete membership privacy breach is possible with Log-Loss scores.
Unified framework analyzes privacy risks from gradients in distributed learning.
problem Analyzing inference privacy risks from gradients in machine learning.
method Unified game-based framework for various attacks, including attribute, property, distributional, and user disclosures.
result Demonstrates inefficacy of data aggregation for privacy against inference attacks.
Proposes a new privacy notion for membership inference attacks on machine learning models.
problem Membership inference attacks on machine learning models.
method Introduces f-Membership Inference Privacy (f-MIP) and μ-Gaussian Membership Inference Privacy (μ-GMIP) to quantify and mitigate privacy risks. result Analyzes likelihood ratio-based attacks and derives μ-GMIP guarantees for stochastic gradient descent (SGD) models. Synthetic data mimics real-world demographics for fairness testing.
problem Lack of complete, representative datasets for fairness testing.
method Construct synthetic datasets using overlapping real and separate datasets.
result Synthetic data yields consistent fairness metrics with real data.
FedIRT enables privacy-preserving psychometric estimation without centralizing data.
problem Privacy and data governance concerns in centralized IRT estimation.
method Federated Item Response Theory (FedIRT) and FedIRT-DP for differentially private estimation.
result FedIRT matches accuracy of centralized estimators while preserving privacy.
Unlike other industries in which intellectual property is patentable, the financial industry relies on trade secrecy to protect its business processes and methods, which can obscure critical financial risk exposures from regulators and the public. We develop methods for sharing and aggregating such risk exposures that …
Study examines auditing fairness in evolving models, identifying strategic updates that preserve audit properties.
problem Auditing fairness in machine learning models that adapt to changing environments.
method Characterizes strategic updates that preserve audit properties, proposes a generic PAC auditing framework.
result Establishes distribution-free auditing bounds for statistical parity using the SP dimension.
Neural networks help auditors efficiently assess financial statements by learning underlying data patterns.
problem Efficiently auditing large volumes of financial statements and journal entries.
method Vector Quantised-Variational Autoencoder (VQ-VAE) neural networks.
result VQ-VAE neural networks can learn a quantized representation of accounting data, uncovering latent factors and providing a representative audit sample.
This review examines deep learning in financial fraud detection over 5 years.
problem Improving deep learning techniques for financial fraud detection.
method Systematic literature review of 57 studies using performance metrics.
result Deep learning models enhance fraud detection across various financial domains.
Proposes practical kernel tests for f-divergences with theoretical guarantees.
problem Two-sample testing and machine unlearning evaluation.
method Regularized f-divergence kernel tests, adaptive to hyperparameters. result Different f-divergences highlight localized differences. Signed compression progress on a sealed audit is goodhart-resistant.
problem Intrinsic motivation for agents to improve their world models by compressing experience.
method Rewarding agents for the signed decrease of a fixed sealed-audit loss.
result Cumulative reward telescopes exactly to endpoint audit improvement, preventing infinite reward push while true audit performance stagnates.
DAISYnt evaluates synthetic data quality and privacy in regulated domains.
problem Balancing data quality and privacy in regulated domains.
method Developed a suite of advanced tests (DAISYnt) to evaluate synthetic data quality and privacy.
result DAISYnt sets a de facto standard for synthetic data evaluation in regulated domains.
Study shows membership inference requires many more samples than training data.
problem Membership inference attacks require more reference samples than training data.
method Investigated the sample complexity for membership inference in Gaussian mean estimation.
result Ω(n + n^2 ρ^2) samples are necessary for successful membership inference attacks.
This paper quantifies how hard it is to identify specific data points in machine learning models.
problem Quantifying the difficulty of identifying specific data points in machine learning models.
method Characterizing optimal attacks and privacy defences, deriving impacts of noise and misspecification, and proposing a new covariance attack.
result The Mahalanobis distance explains the hardness of fixed-target membership inference attacks.
This study improves audit sampling by using sequential procedures with statistical guarantees.
problem Improving audit efficiency and reliability with statistical methods.
method Formulated as a sequential testing problem, defining null and alternative hypotheses, stopping and decision rules, and exact boundary conditions.
result Exact design yields ex ante control of decision error probabilities, and simulation-based implementation approximates this design.
RESHAPE explains financial statement anomalies by aggregating explanations from AENNs.
problem Detecting and explaining accounting anomalies in financial audits is challenging.
method Proposes RESHAPE to explain model output on an aggregated attribute-level.
result RESHAPE provides more comprehensible explanations compared to existing methods.
AI-driven framework improves enterprise financial audits and risk identification.
problem Manual auditing is inefficient and limited by data complexity and evolving fraud tactics.
method Machine learning algorithms (SVM, RF, KNN) applied to a dataset of audit project counts, violations, and fraud instances.
result Random Forest achieves best performance with F1-score of 0.9012, identifying fraud and compliance anomalies.
Fairness audits fail under missing protected labels, especially at zero access.
problem Understanding the reliability of fairness audits with incomplete protected-label data.
method Introduced a seed-calibrated stress test to separate missingness effects from seed-to-seed movement.
result Missing protected labels do not significantly alter fairness mitigation methods, but they can lead to harmful intersectional outcomes.
Audited Conformal Prediction improves conditional coverage in pretrained models under distribution shift.
problem Uncertainty quantification for pretrained models under unknown distribution shift
method Leverages a small labeled dataset to train an audit model for marginal coverage, integrates outputs into conformal prediction framework
result Significantly higher conditional coverage than existing approaches
To help enforce data-protection regulations such as GDPR and detect unauthorized uses of personal data, we develop a new \emph{model auditing} technique that helps users check if their data was used to train a machine learning model. We focus on auditing deep-learning models that generate natural-language text, includi…
Study efficient auditing of ML fairness models.
problem Scalability of auditing ML models for fairness.
method Query-based auditing algorithms for estimating demographic parity.
result Optimal deterministic and practical randomized algorithms for fairness estimation.
Study cost-effective fairness audits with partial feedback, improving over random exploration.
problem Auditing fairness of classifiers with limited true labels.
method Introduces cost model, proposes near-optimal algorithms for black-box and mixture models.
result Significantly lower audit costs compared to natural baselines.
Develops tools to audit ML models for bias and unfairness.
problem Auditing ML models for individual bias and unfairness.
method Formalizes the task as an optimization problem and develops inferential tools for the optimal value.
result Demonstrates the utility of tools in revealing biases in COMPAS recidivism prediction instrument.
Algorithm identifies best arm with biased proxy and selective ground truth audits.
problem Fixed-confidence best-arm identification with biased proxy and selective ground truth.
method Propensity-weighted estimator and adaptive auditing algorithm.
result Plug-in Neyman rule achieves near-oracle audit efficiency.
Proposes PA-DSL for correcting noisy human labels in automated data labeling.
problem Noisy human labels in automated data labeling.
method Uses adjudicated cases to correct noisy human labels and debias analyses.
result Maintains nominal coverage and reduces RMSE by 10-17% relative to using only adjudicated labels.
Motivated by the need to audit complex and black box models, there has been extensive research on quantifying how data features influence model predictions. Feature influence can be direct (a direct influence on model outcomes) and indirect (model outcomes are influenced via proxy features). Feature influence can also …
The rapid growth of text data has motivated the development of machine-learning based automatic text summarization strategies that concisely capture the essential ideas in a larger text. This study aimed to devise an extractive summarization method for A-133 Single Audits, which assess if recipients of federal grants a…