Search NASA⌕ Search

SEARCH · Search NASA

Results for “Distributed training”

Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 577 records · Page 32

Deep Learning At Depth: Estimating subsurface parameters from geophysical monitoring data

Geophysical imaging techniques are a non-invasive way to image the subsurface and understand both subsurface solid (rock/soil) and fluid property distributions and their evolution in time. Inversions of the geophysical data, such as Electrical Resistance Tomography (ERT) data, are solved to estimate the subsurface property distributions, such as conductivity, and many inversion techniques smooth out sharp gradients in rock or fluid property distributions. Sharp gradients in subsurface properties tend to be present in situations with complex subsurface structures, which are common in many subsurface applications. We have successfully demonstrated that it is possible to inform, or constrain, inversions with neural networks trained on synthetic data with complex subsurface structures. Initial results suggest this process may be optimizable to yield property distributions that better represent the true property distributions than the same inversion process without the neural network constraint. Future work would optimize the neural network performance for this application and then apply the synthetic-data trained neural network to real data to understand the utility and performance of this technique for real data sets.

47 OTHER INSTRUMENTATION↗

Convex Q-Learning in Continuous Time with Application to Dispatch of Distributed Energy Resources

Convex Q-learning is a recent approach to reinforcement learning, motivated by the possibility of a firmer theory for convergence, and the possibility of making use of greater a priori knowledge regarding policy or value function structure. This paper explores algorithm design in the continuous time domain, with a finite-horizon optimal control objective. The main contributions are (i) The new Q-ODE: a model-free characterization of the Hamilton-Jacobi-Bellman equation. (ii) A formulation of Convex Q-learning that avoids approximations appearing in prior work. The Bellman error used in the algorithm is defined by filtered measurements, which is necessary in the presence of measurement noise. (iii) Convex Q-learning with linear function approximation is a convex program. It is shown that the constraint region is bounded, subject to an exploration condition on the training input. (iv) The theory is illustrated in application to resource allocation for distributed energy resources, for which the theory is ideally suited.

Lu, Fan↗

A petabyte size electronic library using the N-Gram memory engine

A model library containing petabytes of data is proposed by Triada, Ltd., Ann Arbor, Michigan. The library uses the newly patented N-Gram Memory Engine (Neurex), for storage, compression, and retrieval. Neurex splits data into two parts: a hierarchical network of associative memories that store 'information' from data and a permutation operator that preserves sequence. Neurex is expected to offer four advantages in mass storage systems. Neurex representations are dense, fully reversible, hence less expensive to store. Neurex becomes exponentially more stable with increasing data flow; thus its contents and the inverting algorithm may be mass produced for low cost distribution. Only a small permutation operator would be recalled from the library to recover data. Neurex may be enhanced to recall patterns using a partial pattern. Neurex nodes are measures of their pattern. Researchers might use nodes in statistical models to avoid costly sorting and counting procedures. Neurex subsumes a theory of learning and memory that the author believes extends information theory. Its first axiom is a symmetry principle: learning creates memory and memory evidences learning. The theory treats an information store that evolves from a null state to stationarity. A Neurex extracts information data without a priori knowledge; i.e., unlike neural networks, neither feedback nor training is required. The model consists of an energetically conservative field of uniformly distributed events with variable spatial and temporal scale, and an observer walking randomly through this field. A bank of band limited transducers (an 'eye'), each transducer in a bank being tuned to a sub-band, outputs signals upon registering events. Output signals are 'observed' by another transducer bank (a mid-brain), except the band limit of the second bank is narrower than the band limit of the first bank. The banks are arrayed as n 'levels' or 'time domains, td.' The banks are the hierarchical network (a cortex) and transducers are (associative) memories. A model Neurex was built and studied. Data were 50 MB to 10 GB samples of text, data base, and images: black/white, grey scale, and high resolution in several spectral bands. Memories at td, S(m(sub td)), were plotted against outputs of memories at td-1. S(m(sub td)) was Boltzman distributed, and memory frequencies exhibited self-organized criticality (SOC); i.e., 'l/f(sup beta)' after long exposures to data. Whereas output signals from level n may be encoded with B(sub output) = O(-log(2)f(sup beta)) bits, and input data encoded with B(sub input) = O((S(td)/S(td-1))(sup n)), B(sup output)/B(sub input) is much less than 1 always, the Neurex determines a canonical code for data and it is a lossless data compressor. Further tests are underway to confirm these results with more data types and larger samples.

Bugajski, Joseph M.↗

The use of digital thread for reconstruction of local fiber orientation in a compression molded pin bracket via deep learning

A deep convolutional neural network (DCNN) was used for microstructure reconstruction using artificial intelligence (MR-AI) by predicting local average fiber orientation distributions (FOD) in a 3D prepreg platelet molded composite (PPMC) pin bracket. To train the MR-AI model, surface strain fields from residual stresses simulated in PPMC plates were used as the input to the DCNN. A training dataset included PPMC plates with various degrees of global fiber alignment, based on the information obtained from high-fidelity flow simulation of a pin bracket. Further, the MR-AI model was then deployed to analyze FOD in the 3D pin bracket by conducting thermo-elastic residual stress analysis. Initially, the MR-AI model was established entirely on the synthetic simulation data. Then, a μCT scan of a physically molded pin bracket was used to create a finite element model that provided data for additional validation of the DCNN model. For the μCT scan finite element pin bracket the MR-AI model predicted the distribution of fiber orientation tensor components with MAE of 0.10 indicating a global prediction error of 10%. For the flow simulated pin bracket, the MR-AI model predicted the distribution of fiber orientation tensor components with a global prediction error of 11%. The MR-AI model showed the ability to predict regions of varying alignment in the base and flange of the pin bracket. The proposed MR-AI methodology allows for rapid prediction of FOD in geometrically complex parts and offers a promising path to detecting unique fiber orientation states in molded components.

42 ENGINEERING↗

Solving high-dimensional inverse problems using amortized likelihood-free inference with noisy and incomplete data

Here, we present a likelihood-free probabilistic inversion method based on normalizing flows for high-dimensional inverse problems. The proposed method is composed of two complementary networks: a summary network for data compression and an inference network for parameter estimation. The summary network encodes raw observations into a fixed-size vector of summary features, while the inference network generates samples of the approximate posterior distribution of the model parameters based on these summary features. The posterior samples are produced in a deep generative fashion by sampling from a latent Gaussian distribution and passing these samples through an invertible transformation. We construct this invertible transformation by sequentially alternating conditional invertible neural network and conditional neural spline flow layers. The summary and inference networks are trained simultaneously. We apply the proposed method to an inversion problem in groundwater hydrology to estimate the posterior distribution of the log-conductivity field conditioned on spatially sparse time-series observations of the system’s hydraulic head responses. The conductivity field is represented with 706 degrees of freedom in the considered problem. Comparison with the likelihood-based iterative ensemble smoother PEST-IES method demonstrates that the proposed method accurately estimates the parameter posterior distribution and the observations’ predictive posterior distribution at a fraction of the inference time of PEST-IES.

conditional invertible neural network↗

Team Training and Retention of Skills Acquired Above Real Time Training on a Flight Simulator

Above Real-Time Training (ARTT) is the training acquired on a real time simulator when it is modified to present events at a faster pace than normal. The experiments related to training of pilots performed by NASA engineers (Kolf in 1973, Hoey in 1976) and others (Guckenberger, Crane and their associates in the nineties) have shown that in comparison with the real time training (RTT), ARTT provides the following benefits: increased rate of skill acquisition, reduced simulator and aircraft training time, and more effective training for emergency procedures. Two sets of experiments have been performed; they are reported in professional conferences and the respective papers are included in this report. The retention of effects of ARTT has been studied in the first set of experiments and the use of ARTT as top-off training has been examined in the second set of experiments. In ARTT, the pace of events was 1.5 times the pace in RTT. In both sets of experiments, university students were trained to perform an aerial gunnery task. The training unit was equipped with a joystick and a throttle. The student acted as a nose gunner in a hypothetical two place attack aircraft. The flight simulation software was installed on a Universal Distributed Interactive Simulator platform supplied by ECC International of Orlando, Florida. In the first set of experiments, two training programs RTT or ART7 were used. Students were then tested in real time on more demanding scenarios: either immediately after training or two days later. The effects of ARTT did not decrease over a two day retention interval and ARTT was more time efficient than real time training. Therefore, equal test performance could be achieved with less clock-time spent in the simulator. In the second set of experiments three training programs RTT or ARTT or RARTT, were used. In RTT, students received 36 minutes of real time training. In ARTT, students received 36 minutes of above real time training. In RARTT, students received 18 minutes of real time training and 18 minutes of above real time training as top-off training. Students were then tested in real time on more demanding scenarios. The use of ARTT as top-off training after RTT offered better training than RTT alone or ARTT alone. It is, however, suggested that a similar experiment be conducted on a relatively more complex task with a larger sample of participants. Within the proposed duration of the research effort, the setting up of experiments and trial runs on using ARTT for team training were also scheduled but they could not be accomplished due to extra ordinary challenges faced in developing the required software configuration. Team training is, however, scheduled in a future study sponsored by NASA at Tuskegee University.

Ali, Syed Friasat↗

Adaptive machine learning for time-varying systems: low dimensional latent space tuning

Machine learning (ML) tools such as encoder-decoder convolutional neural networks (CNN) can represent incredibly complex nonlinear functions which map between combinations of images and scalars. For example, CNNs can be used to map combinations of accelerator parameters and images which are 2D projections of the 6D phase space distributions of charged particle beams as they are transported between various particle accelerator locations. Despite their strengths, applying ML to time-varying systems, or systems with shifting distributions, is an open problem, especially for large systems for which collecting new data for re-training is impractical or interrupts operations. Particle accelerators are one example of large time-varying systems for which collecting detailed training data requires lengthy dedicated beam measurements which may no longer be available during regular operations. We present a novel method of adaptive ML for time-varying systems. Our approach is to map very high (N ≈ 100k) dimensional inputs (a combination of scalar parameters and images) into the low dimensional (N ≈ 2) latent space at the output of the encoder section of an encoder-decoder CNN. We then actively tune the low dimensional latent space-based representation of complex system dynamics by the addition of an adaptively tuned feedback vector directly before the decoder sections builds back up to our image-based high-dimensional phase space density representations. This method allows us to learn correlations within and to quickly tune the characteristics of incredibly large parameter space systems and to track their evolution in real time based on feedback without massive new data sets for re-training. We demonstrate that our method can accurately predict and track the phase space of charged particle beams at various locations in a particle accelerator by adaptively adjusting in real-time while the unknown input beam distribution of the accelerator is changing in shape, charge, and offset and while the RF system of the accelerator itself is also changing in an unpredictable way. For FACET-II we demonstrate that such an approach has the potential to use transverse deflecting cavity and energy spread spectrum beam measurements to accurately predict 2D projections of the 6D phase space of the electron beam at the plasma wakefield acceleration interaction point where such diagnostics are unavailable.

47 OTHER INSTRUMENTATION↗

DiffLense: a conditional diffusion model for super-resolution of gravitational lensing data

Abstract Gravitational lensing data is frequently collected at low resolution due to instrumental limitations and observing conditions. Machine learning-based super-resolution techniques offer a method to enhance the resolution of these images, enabling more precise measurements of lensing effects and a better understanding of the matter distribution in the lensing system. This enhancement can significantly improve our knowledge of the distribution of mass within the lensing galaxy and its environment, as well as the properties of the background source being lensed. Traditional super-resolution techniques typically learn a mapping function from lower-resolution to higher-resolution samples. However, these methods are often constrained by their dependence on optimizing a fixed distance function, which can result in the loss of intricate details crucial for astrophysical analysis. In this work, we introduce DiffLense , a novel super-resolution pipeline based on a conditional diffusion model specifically designed to enhance the resolution of gravitational lensing images obtained from the Hyper Suprime-Cam Subaru Strategic Program (HSC-SSP). Our approach adopts a generative model, leveraging the detailed structural information present in Hubble space telescope (HST) counterparts. The diffusion model, trained to generate HST data, is conditioned on HSC data pre-processed with denoising techniques and thresholding to significantly reduce noise and background interference. This process leads to a more distinct and less overlapping conditional distribution during the model’s training phase. We demonstrate that DiffLense outperforms existing state-of-the-art single-image super-resolution techniques, particularly in retaining the fine details necessary for astrophysical analyses.

Computer Science↗

A neural networks framework for real-time unfolding of neutron spectroscopic data at JET

A determination of fast ion population parameters such as intensity and kinetic temperature is important for fusion reactors. This becomes more challenging with finer time resolution of the measurements, since the limited data in each time slice cause increasing statistical variations in the data. This paper describes a framework using Bayesian-regularized neural networks (NNs) designed for such a task. The method is applied to the TOFOR 2.5 MeV fusion neutron spectrometer at JET. NN training data are generated by random sampling of variables in neutron spectroscopy models. Ranges and probability distributions of the parameters are chosen to match the experimental data. Results have shown good performance both on synthetic and experimental data. The latter was assessed by statistical considerations and by examining the robustness and time consistency of the results. The regularization of the training algorithm allowed for higher time resolutions than simple forward methods. The fast execution time makes this approach suitable for real-time analysis with a time resolution limit in the microsecond time scale.

70 PLASMA PHYSICS AND FUSION TECHNOLOGY↗

Uncertainty quantification of a physics-informed model based on sparse identification of a Thermal Energy Distribution System

Integrated energy systems (IES)s are crucial for enhancing the economy and efficiency of power generation sources (e.g., nuclear energy) necessary to unleash American energy dominance. These systems can be integrated with thermal energy storage (TES) and intermittent renewable energies to optimize overall energy use, peak-load regulation, and demand-side responses. However, the stabilization of energy generation, transport, and utilization introduces operational complexities that exceed the challenges of managing each sub-component individually. Currently, though IESs rely on human operators for efficiency and stability, reducing human error risk and enhancing performance through automation is highly desirable. Recent advances at Idaho National Laboratory have demonstrated successful control of the Thermal Energy Distributed System (TEDS). However, the automatic control system depends on a deterministic Sparse Identification of Nonlinear Dynamics with Control (SINDyC) model, which are trained based on simulation data from physics-based simulations. Because of uncertainties in physics-based simulation, SINDyC model results in large discrepancies against experimental data and cannot be reliably used in automatic control. In this paper, we present an innovative approach to address these discrepancies by quantifying uncertainties and developing a more robust model. We first generated trajectories by using first-principles physics codes to encapsulate the experiment. Next, we trained thousands of models by randomly sampling these trajectories. We then collapsed all those models into one probabilistic SINDyC by fitting a multivariate Gaussian distribution onto the resulting coefficient’s distribution. Despite its simplicity, our approach successfully produced 95% confidence intervals that captured the experimental trajectories. It even did so with a higher probability and better U-pooling score across six of the seven relevant quantities of interest (QoIs), as compared to other classical approaches. In conclusion, ongoing research is focusing on generating new experimental trajectories to validate this approach, and on employing Bayesian calibration to refine parametric uncertainties and guide future model development efforts.

22 - GENERAL STUDIES OF NUCLEAR REACTORS↗

Advanced Orion Optimized Laser System Analysis

Contractor shall perform a complete analysis of the potential of the solid state laser in the very long pulse mode (100 ns pulse width, 10-30 hz rep-rate) and in the very short pulse mode (100 ps pulse width 10-30 hz rep rate) concentrating on the operation of the device in the 'hot-rod' mode, where no active cooling the laser operation is attempted. Contractor's calculations shall be made of the phase aberrations which develop during the repped-pulse train, and the results shall feed into the adaptive optics analyses. The contractor shall devise solutions to work around ORION track issues. A final report shall be furnished to the MSFC COTR including all calculations and analysis of estimates of bulk phase and intensity aberration distribution in the laser output beam as a function of time during the repped-pulse train for both wave forms (high-energy/long-pulse, as well as low-energy/short-pulse). Recommendations shall be made for mitigating the aberrations by laser re-design and/or changes in operating parameters of optical pump sources and/or designs.

Source record↗

Characterization of Temperature Heterogeneity in Utility-Scale Power Plant Boilers by Spatially Distributed Ultrasonic Measurements

In extreme environments, even hardened insertion sensors fail quickly. For such environments, we have developed an ultrasonic (US) method for measuring the spatial distribution of temperatures in solid materials and, specifically, across containments of extreme processes. By deploying US sensors in multiple locations, spatial heterogeneity in temperatures inside harsh environments may be characterized. In this note, we update our progress on testing the developed approach in an industrial setting and its performance in describing the temperature distribution inside a 500 MW (electrical) coalfired utility boiler of an electrical power generation plant. We use waveguides (WG) structured to contain echogenic features that produce a train of echoes in response to an excitation pulse. The time of flight between echoes encodes the information on the temperature distribution in the corresponding segment of the WG, which we reconstruct using its parametrization. Five waveguides were welded to the boiler’s heat exchange surface (waterwall) and different locations at the same boiler elevation and produced reliable US waveforms 18 months after the installation. For several weeks during the latest trial, we performed US measurements in five locations spanning half of the boiler’s 14-meter width. The measurements were acquired during regular operation, including load cycling to adjust for demand and intermittent contributions of renewable power sources, and during boiler shutdown for emergency maintenance and the subsequent restart. The interpretation of US waveforms by signal processing resulted in an accurate estimation of temperature distribution along the waveguides. We captured daily cyclical variations in demand-following load. Measurements performed in multiple locations revealed unexpected temperature variations across the boiler. We conclude that an array of developed sensors can provide responsive and spatial temperature measurements while maintaining functionality despite prolonged exposure to an extreme environment.

Walton, Kenneth↗

Fed-DeepONet: Stochastic Gradient-Based Federated Training of Deep Operator Networks

The Deep Operator Network (DeepONet) framework is a different class of neural network architecture that one trains to learn nonlinear operators, i.e., mappings between infinite-dimensional spaces. Traditionally, DeepONets are trained using a centralized strategy that requires transferring the training data to a centralized location. Such a strategy, however, limits our ability to secure data privacy or use high-performance distributed/parallel computing platforms. To alleviate such limitations, in this paper, we study the federated training of DeepONets for the first time. That is, we develop a framework, which we refer to as Fed-DeepONet, that allows multiple clients to train DeepONets collaboratively under the coordination of a centralized server. To achieve Fed-DeepONets, we propose an efficient stochastic gradient-based algorithm that enables the distributed optimization of the DeepONet parameters by averaging first-order estimates of the DeepONet loss gradient. Then, to accelerate the training convergence of Fed-DeepONets, we propose a moment-enhanced (i.e., adaptive) stochastic gradient-based strategy. Finally, we verify the performance of Fed-DeepONet by learning, for different configurations of the number of clients and fractions of available clients, (i) the solution operator of a gravity pendulum and (ii) the dynamic response of a parametric library of pendulums.

Moya, Christian↗

A data-driven framework for sparsity-enhanced surrogates with arbitrary mutually dependent randomness

The challenge of quantifying uncertainty propagation in real-world systems is rooted in the high-dimensionality of the stochastic input and the frequent lack of explicit knowledge of its probability distribution. Traditional approaches show limitations for such problems, especially when the size of the training data is limited. To address these diculties, we have developed a general framework of constructing surrogate models on spaces of stochastic input with arbitrary probability measure irrespective of the mutual dependencies between individual components of the random inputs and the analytical form. The present Data-driven Sparsity-enhancing Rotation for Arbitrary Randomness (DSRAR) framework includes a data-driven construction of multivariate polynomial basis for arbitrary mutually dependent probability measure and a sparsity enhancement rotation procedure. This sparsity-enhancing rotation method was initially proposed in our previous work [1] for Gaussian density distributions, which may not be feasible for non-Gaussian distributions due to the loss of orthogonality after the rotation. To remedy such diculties, we developed a new data-driven approach to construct orthonormal polynomials for polynomials for arbitrary mutually dependent (amdP) randomness, ensuring the constructed basis maintains the orthogonality/near-orthogonality with respect to the density of the rotated random vector, where directly applying the regular polynomial chaos including arbitrary polynomial chaos (aPC) [2] shows limitations due to the assumption of the mutual independence between the components of the random inputs. The developed DSRAR framework leads to accurate recovery, with only limited training data, of a sparse representation of the target functions. The eectiveness of our method is demonstrated in challenging problems such as PDEs and realistic molecular systems within high-dimensional conformational space (O(10)) where the underlying density is implicitly represented by a large collection of sample data, as well as systems with explicitly given non-Gaussian probabilistic measures.

Lei, Huan↗

Improving multiwell petrophysical interpretation from well logs via machine learning and statistical models

Well-log interpretation estimates in situ rock properties along well trajectory, such as porosity, water saturation, and permeability, to support reserve-volume estimation, production forecasts, and decision making in reservoir development. However, due to measurement errors, variability of well logs caused by multiple measurement vendors, different borehole tools, and nonuniform drilling/borehole conditions, estimations of rock properties with original well logs without proper preprocessing may not be accurate, especially in the context of multiwell estimation. Well-log normalization techniques such as two-point scaling and mean-variance normalization are commonly used to improve the robustness of multiwell rock-property estimation. However, these techniques do not consider the correlation between well logs and require subjective knowledge for their effective implementation. To reduce uncertainties and processing time associated with multiwell rock-property estimation from well logs, we develop discriminative adversarial (DA) and linear constraint models for well-log normalization and rock-property estimation. The DA neural network model developed for well-log normalization and interpretation can perform linear and nonlinear well-log normalization while considering the joint distribution of each well log and rock properties. However, the linear constraint model uses an ensemble of predictions from linear models to constrain well-log normalization and rock-property estimation. We also develop a divergence-based type well identification method to select type (training) wells for a test well based on the statistical similarity of associated well-log distributions instead of the interwell distance. We apply the DA model to perform well-log normalization and prediction of permeability for the Seminole San Andres Unit carbonate reservoir. Compared with the permeability predicted with the classical machine learning model without well-log normalization and models with two-point scaling normalization, the DA model yields the most accurate permeability prediction by decreasing the mean-squared error of permeability prediction by 20%–50%.

Geochemistry & Geophysics↗

Detecting False Data Injection Attacks in Smart Grids: A Semi-Supervised Deep Learning Approach

The dependence on advanced information and communication technology increases the vulnerability in smart grids under cyber-attacks. Recent research on unobservable false data injection attacks (FDIAs) reveals the high risk of secure system operation, since these attacks can bypass current bad data detection mechanisms. To mitigate this risk, this paper proposes a data-driven learning-based algorithm for detecting unobservable FDIAs in distribution systems. We use autoencoders for efficient dimension reduction and feature extraction of measurement datasets. Further, we integrate the autoencoders into an advanced generative adversarial network (GAN) framework, which successfully detects anomalies under FDIAs by capturing the unconformity between abnormal and secure measurements. Also, considering that the datasets collected from practical power systems are partially labeled due to expensive labeling costs and missing labels, the proposed method only requires a few labeled measurement data in addition to unlabeled data for training. Numerical simulations in three-phase unbalanced IEEE 13-bus and 123-bus distribution systems validate the detection accuracy and efficiency of this method.

97 MATHEMATICS AND COMPUTING↗

Using automated machine learning for the upscaling of gross primary productivity

Estimating gross primary productivity (GPP) over space and time is fundamental for understanding the response of the terrestrial biosphere to climate change. Eddy covariance flux towers provide in situ estimates of GPP at the ecosystem scale, but their sparse geographical distribution limits larger-scale inference. Machine learning (ML) techniques have been used to address this problem by extrapolating local GPP measurements over space using satellite remote sensing data. However, the accuracy of the regression model can be affected by uncertainties introduced by model selection, parameterization, and choice of explanatory features, among others. Recent advances in automated ML (AutoML) provide a novel automated way to select and synthesize different ML models. In this work, we explore the potential of AutoML by training three major AutoML frameworks on eddy covariance measurements of GPP at 243 globally distributed sites. We compared their ability to predict GPP and its spatial and temporal variability based on different sets of remote sensing explanatory variables. Explanatory variables from only Moderate Resolution Imaging Spectroradiometer (MODIS) surface reflectance data and photosynthetically active radiation explained over 70 % of the monthly variability in GPP, while satellite-derived proxies for canopy structure, photosynthetic activity, environmental stressors, and meteorological variables from reanalysis (ERA5-Land) further improved the frameworks' predictive ability. We found that the AutoML framework Auto-sklearn consistently outperformed other AutoML frameworks as well as a classical random forest regressor in predicting GPP but with small performance differences, reaching an r 2 of up to 0.75. We deployed the best-performing framework to generate global wall-to-wall maps highlighting GPP patterns in good agreement with satellite-derived reference data. This research benchmarks the application of AutoML in GPP estimation and assesses its potential and limitations in quantifying global photosynthetic activity.

54 ENVIRONMENTAL SCIENCES↗

IntraShuffler: A Privacy Preserving Framework for Heterogeneous DP Federated Learning

Heterogeneous Differential Privacy (HDP) in Federated Learning (FL) allows clients to select individual privacy budgets () according to institutional policies and data sensitivity. In practice, many HDP-FL systems employ -aware server aggregation to improve model utility by re-weighting client updates according to their declared privacy budgets. However, gradient updates in FL retain structural patterns induced by non-independent and identically-distributed (non-IID) data, and these additional signals exposed by -aware aggregation create new opportunities for inference by an honest-but-curious server. In this work, we first show that a server equipped with gradient denoising and surrogate modeling can mount a Privacy Inference Attack that infers distributional attributes of clients and links updates from the same client across training rounds, measured via surrogate inference accuracy and linkage success, under realistic knowledge constraints. The Shuffle-Model has been widely studied as a defense against such inference risks by anonymizing update sources, but it is fundamentally incompatible with HDP-FL -aware aggregation. To address this challenge, we propose IntraShuffler, a middleware defense framework designed for HDP-FL systems. IntraShuffler introduces a privacy-aware shuffling mechanism that groups clients into privacy-compatible buckets and performs parameter-level shuffling within each bucket to disrupt persistent gradient structure while preserving -aware aggregation. Experiments across four different datasets show that IntraShuffler reduces gradient recoverability by over 60% and decreases surrogate inference accuracy from 0.78 to 0.33 while maintaining comparable model utility across multiple FL aggregation rules.

Riya, Farhin Farhad [ORNL]↗