Edgeworth Accountant calculates privacy loss under differential privacy compositions efficiently.
problem Efficiently computing overall privacy loss under composition of private algorithms.
method Analytical approach using f-differential privacy framework and Edgeworth expansion. result Non-asymptotic (ε,δ)-differential privacy bounds with reduced computational cost. Study detects signals in spiked Wigner models using log likelihood ratio.
problem Detecting signals in rank-one spiked Wigner models with non-Gaussian noise.
method Proved asymptotic normality of log likelihood ratio and computed error thresholds.
result Optimal signal-to-noise ratio threshold for reliable detection.
In this work, a deep learning-based method for log-likelihood ratio (LLR) lossy compression and quantization is proposed, with emphasis on a single-input single-output uncorrelated fading communication setting. A deep autoencoder network is trained to compress, quantize and reconstruct the bit log-likelihood ratios cor…
New method for robust distribution alignment using log-likelihood ratio and normalizing flows.
problem Distribution alignment challenges in deep learning.
method Log-likelihood ratio statistic and normalizing flows.
result Minimizing the proposed objective yields robust domain alignment.
In deep neural network, the cross-entropy loss function is commonly used for classification. Minimizing cross-entropy is equivalent to maximizing likelihood under assumptions of uniform feature and class distributions. It belongs to generative training criteria which does not directly discriminate correct class from co…
We propose the method to sanitize the privacy of the IFM(Input Feature Map)s that are fed into the layers of CNN(Convolutional Neural Network)s. The method introduces the degree of the sanitization that makes the application using a CNN be able to control the privacy loss represented as the ratio of the probabilistic a…
SPRT-TANDEM improves sequential classification accuracy with fewer samples.
problem Efficiently classifying sequential data with high accuracy and low sampling cost.
method Deep neural network-based SPRT algorithm that estimates log-likelihood ratio of two hypotheses.
result SPRT-TANDEM achieves statistically significantly better classification accuracy than other classifiers with fewer samples.
In this note, we study the relationship between the variational gap and the variance of the (log) likelihood ratio. We show that the gap can be upper bounded by some form of dispersion measure of the likelihood ratio, which suggests the bias of variational inference can be reduced by making the distribution of the like…
This paper quantifies privacy loss in exploratory data analysis.
problem Privacy loss in exploratory data analysis is often overlooked in privacy budgets.
method Quantitative analysis of privacy loss for statistical functions.
result Privacy loss must be considered in calculating machine learning privacy budgets.
The paper analyzes and proposes methods for privately sharing individual privacy losses using per-instance differential privacy.
problem The standard differential privacy framework provides a worst-case bound that may not accurately reflect individual privacy losses.
method The paper analyzes per-instance differential privacy and proposes methods to privately and accurately publish per-instance privacy losses.
result The methods privately and accurately publish per-instance differential privacy losses with minimal additional privacy cost.
A new method for tighter privacy loss accounting in adaptive analyses.
problem Ensuring individual privacy in adaptive analyses while staying within a privacy budget.
method A personalized privacy loss estimate and a Rényi differential privacy filter.
result Personalized privacy loss accounting can be practical and tighter than existing methods.
EQ-Net combines LLR estimation and quantization using deep learning.
problem Unified solution for LLR estimation and quantization.
method Two-stage algorithm using LLR compression as a pretext task.
result Achieves state-of-the-art results with gains in efficiency and latency.
A new privacy accountant for Gaussian differential privacy measures individual privacy losses.
problem Bounding differential privacy loss for each participant in data analysis.
method Developed a privacy accountant for adaptive compositions of randomised mechanisms using Gaussian differential privacy.
result Provided optimal bounds for the Gaussian mechanism and constructed an approximative individual privacy accountant.
When the in-sample Sharpe ratio is obtained by optimizing over a k-dimensional parameter space, it is a biased estimator for what can be expected on unseen data (out-of-sample). We derive (1) an unbiased estimator adjusting for both sources of bias: noise fit and estimation error. We then show (2) how to use the adjust…
wd1 improves reasoning in dLLMs by optimizing policies without policy ratios.
problem Improving reasoning in diffusion-based large language models through RL.
method wd1: ratio-free policy optimization using weighted log-likelihood.
result wd1 outperforms diffusion-based GRPO while requiring lower computational cost.
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.
A new method detects changes in multivariate data using random forests.
problem Detecting changes in multivariate data.
method A computationally feasible search method using random forests and class probability predictions.
result Consistently locates change points in simulations.
Study on privacy leakage in noisy gradient descent algorithms.
problem Information leakage of iterative randomized learning algorithms about training data.
method Analyzes the dynamics of Rényi differential privacy loss in noisy gradient descent algorithms.
result Privacy loss converges exponentially fast for smooth and strongly convex loss functions.
Paper simplifies DP composition for adaptive privacy budgets, enabling better privacy and accuracy in deep learning.
problem Tension between efficiency and flexibility in DP composition theorems.
method Rényi Differential Privacy (RDP) for adaptive privacy budgets, proving simpler composition theorem with smaller constants.
result Practical DP composition for adaptive privacy budgets, enabling better privacy and accuracy in deep learning.
The paper improves privacy accounting for discrete-valued mechanisms and the subsampled Gaussian mechanism.
problem Improving the accuracy and efficiency of differential privacy accounting for discrete outputs.
method Uses fast Fourier transform (FFT) for rigorous error analysis and accounting of privacy loss.
result Provides strict lower and upper bounds for (ε,δ)-values, demonstrating up to 75% reduction in noise variance. Generative Adversarial Networks (GANs) can achieve state-of-the-art sample quality in generative modelling tasks but suffer from the mode collapse problem. Variational Autoencoders (VAE) on the other hand explicitly maximize a reconstruction-based data log-likelihood forcing it to cover all modes, but suffer from poore…
We consider the problem of aligning a pair of databases with jointly Gaussian features. We consider two algorithms, complete database alignment via MAP estimation among all possible database alignments, and partial alignment via a thresholding approach of log likelihood ratios. We derive conditions on mutual informatio…
The paper optimizes designs for distinguishing between Gaussian process models.
problem Discriminating between two Gaussian process models.
method Sequential and static design criteria, including Kullback Leibler divergences and log-likelihood ratios.
result Necessary conditions for optimal design measures are provided.
Single-microphone, speaker-independent speech separation is normally performed through two steps: (i) separating the specific speech sources, and (ii) determining the best output-label assignment to find the separation error. The second step is the main obstacle in training neural networks for speech separation. Recent…
Paper improves privacy and utility of SGD with bounded domain and smooth losses.
problem Lack of tight privacy bounds and practical assumptions in DPSGD.
method Rigorous privacy characterization for DPSGD with general L-smooth and non-convex loss functions, tracking privacy loss over iterations.
result Privacy loss converges without convexity assumption for bounded domain, improving utility.
We consider the problem of minimizing a convex risk with stochastic subgradients guaranteeing ε-locally differentially private (ε-LDP). While it has been shown that stochastic optimization is possible with ε-LDP via the standard SGD (Song et al., 2013), its convergence rate largely depends on the learning rate, w…
Sequential hypothesis testing is a desirable decision making strategy in any time sensitive scenario. Compared with fixed sample-size testing, sequential testing is capable of achieving identical probability of error requirements using less samples in average. For a binary detection problem, it is well known that for k…
Improved privacy analysis for stochastic gradient descent.
problem Analyzing privacy leakage in noisy stochastic gradient descent.
method Modeling Rényi divergence dynamics with Langevin diffusions, proving exponential privacy loss convergence for smooth and strongly convex objectives.
result Privacy loss converges exponentially fast for smooth and strongly convex objectives under constant step size.
Develops a computationally tractable high-dimensional differential privacy estimator.
problem Differential privacy in high dimensions is computationally intractable.
method Combines high-dimensional robust statistics with differential privacy techniques.
result A computationally tractable algorithm with dimension-independent privacy loss.
Test log-likelihood comparisons can be misleading.
problem Misinterpretation of test log-likelihood in model comparison.
method Simple examples of model comparison and forecast accuracy.
result Test log-likelihood does not always correlate with model accuracy.
FOOD detects out-of-distribution samples quickly without needing OOD data.
problem Detecting out-of-distribution samples efficiently in neural networks.
method Extended DNN classifier with Gaussian layer and log likelihood ratio test.
result FOOD achieves state-of-the-art performance and is fast and applicable.
Differentially private (DP) machine learning has recently become popular. The privacy loss of DP algorithms is commonly reported using (ε,δ)-DP. In this paper, we propose a numerical accountant for evaluating the privacy loss for algorithms with continuous one dimensional output. This accountant can be appl…
Corrects pseudo log-likelihood method issues in various applications.
problem Log-likelihood function unbounded issues in pseudo log-likelihood methods.
method Provided a counterexample and corrected algorithms in previous literature.
result Ensured well-definedness of maximum pseudo log-likelihood estimation.
New framework for DP-SMO with near-optimal privacy-loss trade-off.
problem Optimal trade-off between privacy and population loss in DP-SMO.
method General framework using Phased-ERM method and black-box optimization.
result Near-linear time algorithms with near-optimal guarantees.
Obtaining labels can be costly and time-consuming. Active learning allows a learning algorithm to intelligently query samples to be labeled for efficient learning. Fisher information ratio (FIR) has been used as an objective for selecting queries in active learning. However, little is known about the theory behind the …
This paper sharpens privacy guarantees for high-dimensional PCA under differential privacy.
problem Understanding the exact privacy loss in high-dimensional PCA with differential privacy.
method Analyzes the exponential mechanism in a model-free setting for high-dimensional PCA.
result Sharp utility and privacy characterizations in high dimensions show the difficulty of detecting a target individual's presence.
Constructs models to meet any DP requirement without retraining.
problem Evolving privacy requirements at inference time.
method Two post-processing techniques: random selection and linear combination.
result Final private models satisfying any target privacy parameter.
A new VIS approach improves log-likelihood estimation in latent variable models.
problem Challenges in achieving high log-likelihood with VI for complex posterior distributions.
method Uses forward χ2 divergence to optimize proposal distribution for better log-likelihood estimation. result Consistently outperforms state-of-the-art baselines in log-likelihood and parameter estimation.
New method optimizes clustering with better log-likelihood landscape.
problem Nonconvex log-likelihood optimization in model-based clustering.
method Entropic optimal transport loss for Sinkhorn-EM algorithm.
result New loss function avoids spurious local optima.
Normalization layers improve the accuracy of Differentially Private training of deep neural networks.
problem Reduced accuracy in deep neural networks with Differentially Private training.
method Proposed a novel method for integrating batch normalization with Differentially Private Stochastic Gradient Descent (DPSGD) without additional privacy loss.
result Training deeper networks with better utility-privacy trade-off is possible.
We develop a novel approximate Bayesian computation (ABC) framework, ABCDP, that produces differentially private (DP) and approximate posterior samples. Our framework takes advantage of the Sparse Vector Technique (SVT), widely studied in the differential privacy literature. SVT incurs the privacy cost only when a cond…
New method calculates DMN log-likelihood faster.
problem Precise and fast computation of DMN log-likelihood.
method Derived a closed form expression using gamma function properties.
result Closed form calculation is faster with same accuracy.
Extends dimension reduction to data-driven settings without gradients.
problem Gradient-based dimension reduction limitations in data-driven settings.
method Score ratio matching framework, tailored parameterization, regularization, eigenvalue deflation.
result Outperforms standard score-matching for problems with low-dimensional structure.
This paper develops embeddings that preserve likelihood-based statistical inference.
problem Modern machine learning embeddings destroy the geometric structure required for likelihood-based inference.
method Developed a rigorous theory of likelihood-preserving embeddings and introduced the Likelihood-Ratio Distortion metric.
result Controlling the distortion Δn is necessary and sufficient for preserving inference. We introduce a dynamic credit portfolio framework where optimal investment strategies are robust against misspecifications of the reference credit model. The risk-averse investor models his fear of credit risk misspecification by considering a set of plausible alternatives whose expected log likelihood ratios are penal…
Paper extends FFT-based differential privacy method to heterogeneous compositions.
problem Computing accurate differential privacy guarantees for mixed mechanisms.
method Uses Fast Fourier Transform (FFT) for error analysis and parameter selection.
result Provides tighter bounds for heterogeneous compositions compared to homogeneous cases.
Identifying anomalous patterns in real-world data is essential for understanding where, when, and how systems deviate from their expected dynamics. Yet methods that separately consider the anomalousness of each individual data point have low detection power for subtle, emerging irregularities. Additionally, recent dete…
In this work, a deep learning-based quantization scheme for log-likelihood ratio (L-value) storage is introduced. We analyze the dependency between the average magnitude of different L-values from the same quadrature amplitude modulation (QAM) symbol and show they follow a consistent ordering. Based on this we design a…