Optimizes sparse fine-tuning for privacy in neural networks.
problem Performance gap between DP-SGD and non-private fine-tuning.
method Optimization-based approach using private gradient information for selecting trainable weights.
result Our selection method leads to better prediction accuracy compared to existing approaches.
DPZero fine-tunes large models privately without backpropagation.
problem Memory and privacy challenges in fine-tuning large language models.
method DPZero uses zeroth-order methods for private fine-tuning, avoiding backpropagation.
result DPZero achieves private fine-tuning of RoBERTa and OPT on various tasks.
New algorithms improve privacy and utility of large language models.
problem Privacy-preserving fine-tuning of large language models.
method Meta-framework for differentially private fine-tuning, inspired by recent success in fine-tuning.
result Private fine-tuned models achieve utility close to non-private models, with improved privacy and efficiency.
Private learning can perform well in high dimensions, contrary to known results.
problem When does differentially private learning not suffer in high dimensions?
method Introduced a condition called restricted Lipschitz continuity to derive improved bounds for excess empirical and population risks.
result Gradients in private fine-tuning of large models are mostly controlled by a few principal components, similar to conditions for convex settings.
Improved differentially private deep learning with group-wise clipping techniques.
problem Efficiency and privacy trade-offs in deep learning models.
method Group-wise clipping techniques (per-layer and per-device) to reduce compute time and memory overhead.
result Private learning with group-wise clipping achieves similar or better performance than non-private learning with less wall time.
Broad adoption of machine learning techniques has increased privacy concerns for models trained on sensitive data such as medical records. Existing techniques for training differentially private (DP) models give rigorous privacy guarantees, but applying these techniques to neural networks can severely degrade model per…
Differentially private learning avoids iterative optimization in parameter space.
problem Differentially private training of neural networks
method Hypernetworks trained on public datasets
result Significantly reduces noise in parameter space
Diffusion models generate private synthetic images with high quality.
problem Generating private synthetic images from sensitive datasets.
method Differentially private fine-tuning of pre-trained diffusion models.
result Improved synthetic data quality and accuracy compared to non-private models.
Improved DMs with DP-SGD for generating private images.
problem Low privacy of DMs and lack of good privacy-utility tradeoff.
method Adopt LDMs with DP-SGD on attention modules of LDMs.
result Generates high-quality DP images with reduced parameters.
FMP improves personalized ASR models on private devices.
problem Limited accuracy and privacy in federated fine-tuning of ASR models.
method FMP estimates global and personalized marginal distributions and adjusts NNLM probabilities.
result FMP achieves modest WER reductions on ASR rescoring tasks.
Improves zeroth-order optimization for private machine learning with public data.
problem High computation and memory cost of first-order DP methods.
method PAZO (Public Data Assisted Zeroth-order Optimization) framework.
result Achieves superior privacy/utility tradeoffs across tasks.
DiSK improves DP optimizers by simplifying Kalman filtering for better performance.
problem Performance drop of DP optimizers in large-scale training due to noise injection.
method DiSK uses Kalman filtering to denoise privatized gradients and refine gradient estimations.
result DiSK achieves significant performance improvements over standard DP optimizers in large-scale training.
Bayesian optimization is a powerful tool for fine-tuning the hyper-parameters of a wide variety of machine learning models. The success of machine learning has led practitioners in diverse real-world settings to learn classifiers for practical problems. As machine learning becomes commonplace, Bayesian optimization bec…
Privacy-preserving reinforcement learning from human feedback using decoupled reward modeling.
problem Training large language models with sensitive user information while preserving privacy.
method Proposes a privacy-preserving framework that imposes differential privacy on reward learning only.
result Privacy contributes an additional additive term to the suboptimality gap, and the upper bound is rate-optimal up to logarithmic factors.
Recent state-of-the-art language models utilize a two-phase training procedure comprised of (i) unsupervised pre-training on unlabeled text, and (ii) fine-tuning for a specific supervised task. More recently, many studies have been focused on trying to improve these models by enhancing the pre-training phase, either vi…
Zeroth-order optimization methods lack inherent privacy guarantees.
problem Ensuring differential privacy in zeroth-order optimization methods.
method Analyzing ZO-GD with and without random initialization for convex and strongly convex objectives.
result ZO-GD is not differentially private for strongly convex objectives and can have superlinear privacy loss.
LoRA and privacy: Random projections help but not always.
problem Ensuring differential privacy in LoRA fine-tuning.
method Wishart projection mechanism and noisy variants.
result LoRA is not inherently private, but low-rank fine-tuning can be more private.
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.
Few-shot DP image classification models need more data as privacy increases.
problem Few-shot DP image classification challenges in personalization and federated learning.
method Exhaustive experiments on various parameters affecting few-shot DP accuracy and vulnerability.
result Increasing shots per class is necessary for DP accuracy as privacy increases.
Improved DP-SGD on large models achieves high accuracy on image classification tasks.
problem Differentially private image classification often degrades performance.
method Careful hyper-parameter tuning and signal propagation techniques.
result Achieved 81.4% top-1 accuracy on CIFAR-10 under (8, 10^{-5})-DP.
SOTERIA optimizes neural networks for secure inference with minimal overhead.
problem Protecting user privacy in ML-as-a-service models with low overhead.
method Neural architecture search with dual objectives of accuracy and cryptographic efficiency.
result SOTERIA constructs efficient models for secure inference.
Study explains how noisyGD with DP improves feature learning despite high dimensionality.
problem Improving feature learning in differential privacy settings with noisyGD.
method Layer-peeled model in representation learning, error bound analysis, feature normalization, PCA.
result Misclassification error is independent of dimension in NC, and PCA improves testing accuracy.
New mechanism protects neural network weights from privacy attacks during self-supervised learning.
problem Privacy risks during fine-tuning stage of self-supervised learning.
method Proposes a novel differential privacy mechanism using additive logistic noise.
result Reduces membership inference attack accuracy to 50% while maintaining below 5% performance loss.
New measure FTC quantifies how much a ReLU network can fine-tune.
problem Analyzing memorization capacity in fine-tuned neural networks.
method Defined Fine-Tuning Capacity (FTC) for additive fine-tuning of ReLU networks.
result Upper and lower bounds on FTC for 2 and 3-layer ReLU networks.
This paper studies trade-offs in private prediction methods.
problem Leakage of training data information in machine learning predictions.
method Private training and private prediction methods with trade-offs.
result Private training methods outperform private prediction methods in various settings.
Private method measures nonlinear correlations between data hosted across two entities.
problem Measuring nonlinear correlations between sensitive data hosted across multiple parties while preserving privacy.
method Differentially private estimator of distance correlation.
result First private estimator of nonlinear correlations in a multi-party setup.
This paper re-evaluates hyperparameters for fine-tuning pre-trained models.
problem Current hyperparameter settings for fine-tuning are often ad-hoc and fixed.
method Empirical evaluation of learning rate, batch size, and momentum for fine-tuning.
result Optimal hyperparameters are not only dataset-dependent but also sensitive to domain similarity.
Boosted ensemble of decision tree (DT) classifiers are extremely popular in international competitions, yet to our knowledge nothing is formally known on how to make them \textit{also} differential private (DP), up to the point that random forests currently reign supreme in the DP stage. Our paper starts with the proof…
New method quantifies uncertainty in fine-tuned LLMs using LoRA ensembles.
problem Uncertainty in fine-tuned LLMs and how to trust their predictions.
method Posterior approximations using low-rank adaptation ensembles.
result Unexpected retention of acquired knowledge during fine-tuning in overfitting regime.
BERT fine-tuning is unstable due to optimization issues, not forgetting or dataset size.
problem Stability of fine-tuning BERT-based models across different random seeds.
method Analysis of BERT, RoBERTa, and ALBERT fine-tuned on GLUE datasets, identifying optimization difficulties as the cause of instability.
result Fine-tuning instability is due to optimization difficulties leading to vanishing gradients, not forgetting or dataset size.
New algorithms for privately learning decision lists and halfspaces.
problem Private learning of decision lists and halfspaces.
method Differentially private algorithms for PAC and online models.
result Private algorithms match or surpass non-private guarantees.
Private PGB boosts synthetic data quality using GANs and privacy techniques.
problem Differentially private GANs struggle with convergence and poor output quality.
method Combines reweighted samples from GAN training using Private Multiplicative Weights method.
result Improves synthetic data quality across various datasets and tasks.
Improved code translation by preserving structure with composed fine-tuning.
problem Improving code translation accuracy with unlabeled code outputs.
method Pre-trained denoiser to capture output structure, composed fine-tuning to fine-tune predictor.
result Composed fine-tuning significantly improves generalization over standard fine-tuning.
Near-optimal private tests for simple and MLR hypotheses developed under Gaussian differential privacy.
problem Developing private tests for simple and MLR hypotheses under Gaussian differential privacy.
method A private mean estimator with data-driven clamping bounds, constructing private test statistics.
result Private tests achieve the same asymptotic relative efficiency as non-private most powerful tests.
The paper introduces a Hessian-based method to improve generalization in fine-tuned deep neural networks.
problem Improving generalization in fine-tuned deep neural networks, especially in noisy conditions.
method PAC-Bayesian analysis to identify a Hessian-based distance measure, proving generalization bounds, and developing an algorithm with a generalization error guarantee.
result Hessian-based distance measure correlates well with observed generalization gaps and can match the scale of these gaps in practice.
PEARL uses AI to replicate private equity performance with liquid assets.
problem Lack of access to private equity due to high costs and complexity.
method Combines AI with liquid assets, incorporating asymmetry for better performance.
result Model outperforms liquid proxies and aligns with private equity benchmarks.
Fine-tuning LLMs improves capability but harms safety, study finds.
problem Balancing capability and safety in LLM fine-tuning.
method Theoretical framework and numerical experiments for two safety-aware fine-tuning strategies.
result Characterization of fundamental limits of safety-capability trade-off in LLM fine-tuning.
Fine-tuning harms in-context learning, but restricting updates to the value matrix improves zero-shot performance.
problem Fine-tuning harms in-context learning, reducing zero-shot performance on unseen tasks.
method Theoretical analysis of linear attention models, identifying conditions for degraded few-shot performance.
result Restricting updates to the value matrix improves zero-shot performance while preserving in-context learning.
We present a provably optimal differentially private algorithm for the stochastic multi-arm bandit problem, as opposed to the private analogue of the UCB-algorithm [Mishra and Thakurta, 2015; Tossou and Dimitrakakis, 2016] which doesn't meet the recently discovered lower-bound of Ω(εKlog(T)) [Shar…
Public pretraining improves private model training even in extreme distribution shift scenarios.
problem Improving private model training accuracy in settings with large distribution shift.
method Empirical evaluation and theoretical explanation of public representations improving private training accuracy.
result Public representations can improve private training accuracy by up to 67% over private training from scratch in settings with large distribution shift.
Measures consistency of tabular LLM predictions under fine-tuning multiplicity.
problem Conflicting predictions from fine-tuned tabular LLMs.
method Local stability measure in embedding space.
result Probabilistic guarantees on prediction consistency under multiplicity.
Private learning of Gaussian Mixture Models without boundedness assumptions.
problem Private estimation of parameters of Gaussian Mixture Models with unbounded components.
method Reduction to non-private problem, blackbox privatization, Moitra and Valiant's algorithm.
result First sample complexity upper bound and polynomial time algorithm for privately learning GMMs.
Self-play fine-tuning improves diffusion models for text-to-image generation.
problem Plateauing performance of diffusion models after data saturation.
method Self-play fine-tuning (SPIN-Diffusion) using competition among model versions.
result Significantly improved model performance and human preference alignment.
The paper develops a theory linking pretraining and fine-tuning in neural networks.
problem Understanding how initialization choices impact feature learning and generalization in neural networks.
method Analytical theory of diagonal linear networks, deriving generalization error as a function of initialization parameters and task statistics.
result Different initialization choices place networks into four fine-tuning regimes with varying abilities to support feature learning and generalization.
Private estimation with public data reduces sample complexity.
problem Estimating private distributions with limited public data.
method Differentially private estimation with public data under constraints of pure or concentrated DP.
result Public data can significantly reduce private sample complexity for estimation.
New method reduces fine-tuning cost for reused models.
problem Repeating fine-tuning costs with outdated foundation models.
method Portable Reward Tuning (PRT) trains a reward model to maximize the same loss function as fine-tuning.
result PRT achieves comparable accuracy to inference-time tuning with less inference cost.
Compact models match or exceed GPT's performance in financial news sentiment analysis.
problem Improving financial sentiment analysis models without large computational costs.
method Fine-tuning non-generative, small-sized models (FinBERT, FinDRoBERTa) on a novel market score database.
result Fine-tuned models outperform GPT-3.5 and GPT-4 in zero-shot learning for financial news sentiment analysis.
Develops Merton's model for private companies using DDM.
problem Lack of observable asset values for private companies.
method Uses dividend discount model (DDM) to develop structural model.
result Obtains closed-form formulas for equity and liability values, default probability.