Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Continual Release Moment Estimation with Differential Privacy
Authors: Nikita Kalinin, Jalaj Upadhyay, Christoph H. Lampert
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate JME s effectiveness in two applications: estimating the running mean and covariance matrix for Gaussian density estimation and model training with DP-Adam. (Abstract) 5 Experiments Our main contributions in this work are both algorithmic and theoretical. Specifically, JME is the general purpose technique for moment estimation, which is promising for some scenarios and less promising for others. However, it is also a practical algorithm that can be easily implemented and integrated into standard machine learning pipelines. To demonstrate this, we report on the experimental result of using Algorithm 1 and Algorithm 2 in two exemplary settings, reflecting the application scenarios described above. Private Gaussian density estimation. From a given data distribution, p(x) = N(µ, Σ), we sample n = 200 data points and use either JME or PP to form a private estimates, bpt(x) = N(bµt, bΣt), at each step t = 1, . . . , n, of the continuous release process. Private model training with Adam. We train a convolutional network on the CIFAR-10 dataset with DP-Adam, which is privatized either with JME or PP. |
| Researcher Affiliation | Academia | Nikita P. Kalinin Institute of Science and Technology Austria (ISTA) Klosterneuburg, Austria EMAIL Jalaj Upadhyay Department of Computer Science Rutgers University Piscataway, NJ 08854, USA EMAIL Christoph H. Lampert Institute of Science and Technology Austria (ISTA) Klosterneuburg, Austria EMAIL |
| Pseudocode | Yes | Algorithm 1 Joint Moment Estimation (JME) Algorithm 2 Differentially Private JME Adam |
| Open Source Code | No | Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [No] , Justification: We provide pseudocode for all the algorithms we use and believe they are easily reproducible. However, we do not publish any source code at this time. |
| Open Datasets | Yes | We train a convolutional network on the CIFAR-10 dataset with DP-Adam, which is privatized either with JME or PP. |
| Dataset Splits | No | From a given data distribution, p(x) = N(µ, Σ), we sample n = 200 data points and use either JME or PP to form a private estimates... We train a convolutional network on the CIFAR-10 dataset with DP-Adam... |
| Hardware Specification | No | Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? Answer: [No] Justification: All our experiments require minimal resources and can be run on a single computer or, e.g., in Google Colab. |
| Software Dependencies | No | The paper mentions "Adam optimizer" and "Batch Norm [23]" as components, but does not specify any software libraries or their versions (e.g., Python version, specific deep learning framework like PyTorch or TensorFlow version) used for implementation. |
| Experiment Setup | Yes | Table 3: Hyperparameters for CIFAR-10 Experiments. Medium-privacy experiments use a batch size of 256, compared to 1 in the high-privacy regime, while using a noise multiplier of σε,δ = 1 for the medium-privacy regime and σε,δ = 2 for the high-privacy regime. JME and joint clipping require an additional hyperparameter scaling which is optimized to find the best value for those runs. We also find it helpful to clip the updates; for this, we use the same clipping norm. Method Epoch 1 Epoch 2 ... Epoch 10 DP-Adam-JME 12.43 ... 30.38 DP-Adam-Clip 13.54 ... 28.14 DP-Adam-Debiased 12.41 ... 29.14 DP-Adam 10.29 ... 10.48 Table 3: Hyperparameters for CIFAR-10 Experiments. Medium-privacy experiments use a batch size of 256, compared to 1 in the high-privacy regime, while using a noise multiplier of σε,δ = 1 for the medium-privacy regime and σε,δ = 2 for the high-privacy regime. JME and joint clipping require an additional hyperparameter scaling which is optimized to find the best value for those runs. We also find it helpful to clip the updates; for this, we use the same clipping norm. |