Saturday, November 21, 2026
Science
No Result
View All Result
  • Login
  • HOME
  • SCIENCE NEWS
  • CONTACT US
  • HOME
  • SCIENCE NEWS
  • CONTACT US
No Result
View All Result
Scienmag
No Result
View All Result
Home Science News

Tabular Federated Learning Updates Expose Private Records Under Gradient Inversion Attacks

November 21, 2026
in Science News
Veronica Carney
By Veronica Carney Scienmag Editorial Profile - Federated Learning
Reading Time: 6 mins read
0
Tabular Federated Learning Updates Expose Private Records Under Gradient Inversion Attacks

Tabular Federated Learning Updates Expose Private Records Under Gradient Inversion Attacks

65
SHARES
587
VIEWS
Share on FacebookShare on Twitter
ADVERTISEMENT

In distributed machine learning settings where participants collaborate without centralizing raw records, shared training updates can reveal private tabular information under gradient inversion attacks. A systematic study published in the International Journal of Information Security evaluated how honest-but-curious coordinating servers can reconstruct client training records from transmitted gradients in FedSGD and model deltas in FedAvg. The evaluation established that while federated learning keeps underlying datasets local, communicated mathematical updates expose individual records, particularly when client batches are small or updates summarize few distinct entries. Across diverse configurations, model architecture and client batch size acted as primary levers controlling the degree of reconstructable information, demonstrating that noncryptographic federated protocols cannot be treated as inherent guarantees of data privacy.

The investigation framed the security analysis around a centralized federated learning setting governed by an honest-but-curious server. Under this noncryptographic threat model, the coordinating server faithfully adheres to the prescribed training algorithms and distributes model weights, but actively inspects received client updates to infer the private records that generated them. In the FedSGD protocol, the server observes raw gradients computed across individual client mini-batches, whereas in FedAvg, it observes parameter deltas produced after multiple local optimization steps. The attacker possessed comprehensive knowledge of the public model architecture, loss functions, learning rates, aggregation rules, and public input schemas, including categorical cardinalities and numerical feature positions. To establish an upper bound on feature recoverability, the main experiments provided the attacker with true batch targets, while sensitivity analyses evaluated label-unknown scenarios using dummy targets. Reconstruction was performed exclusively by matching simulated updates to observed updates without granting the server access to local normalization parameters, client marginal distributions, or raw data stores.

To execute gradient inversion on tabular records, the authors adapted the optimization-based gradient matching pipeline from the open-source LeakPro framework. For each intercepted client update, the adversary initialized candidate records from Gaussian noise in the feature space and iteratively minimized a cosine distance objective between the observed update and the update simulated from reconstructed candidates. The optimization executed across 10,000 iterations using the Adam optimizer with an initial learning rate of 0.06 and linear decay. Because tabular records comprise heterogeneous mixtures of continuous measurements and discrete categorical variables, the researchers implemented Hungarian matching on encoded feature tensors to align reconstructed rows with ground truth records before measuring accuracy. Feature-level recovery was evaluated using TabLeak accuracy, where numerical features were considered correct if recovered within 0.319 local standard deviations of the true value and categorical attributes required exact class matches. Crucially, the researchers also calculated exact match rates to measure complete row recovery, alongside comparisons against uniform random guesses and client marginal prior baselines.

Recognizing that comparing federated models at matched communication round counts introduces severe confounding, the authors introduced an exposure-aligned evaluation protocol. In standard federated benchmarking, round counts mask substantial differences in the actual volume of data processed when local batch sizes or local epoch numbers vary. The research team instead measured training progress through effective client exposure milestones, defining exposure as the cumulative number of training examples processed divided by the effective client dataset size. Training runs continued until the least exposed client attained a prescribed exposure threshold, fixing exposure at 25 passes for FedSGD and 24 passes for FedAvg local computation experiments. The empirical evaluation evaluated four tabular datasets: the binary classification Adult dataset spanning 14 features, California Housing regression comprising eight continuous features, a private multiclass classification benchmark containing 101 features, and the MIMIC-IV clinical dataset modeling in-hospital mortality with a positive class weight of 13 to handle acute imbalance. Centralized data splits and client partitions enabled repeatable evaluation across multiple random seeds.

Across the benchmark datasets, the FedSGD experiments demonstrated that update aggregation strongly dictates reconstruction vulnerability, though task formulation modulates the response. For one-hot multilayer perceptrons and one-hot ResNet architectures, smaller client batch sizes consistently yielded the highest reconstruction accuracy. On Adult and the private multiclass benchmark, reconstruction degraded sharply as client batch sizes expanded from one up to 256 records. In contrast, the California Housing regression task revealed a distinct post-training pattern: at the final attack point, batch size one updates exhibited lower reconstruction than intermediate batches, with ResNet reconstruction rising from 0.371 at batch size one to 0.748 at batch size two, and the small multilayer perceptron increasing from 0.462 to 0.815. However, this anomaly did not denote generalized safety for unit-sized regression batches, as initialized attack points before training produced near-total recovery of 0.992 for ResNet and 0.967 for the multilayer perceptron. Across all benchmarks, model training attenuated initial leakage levels without eliminating reconstructability at small batch granularities.

The clinical prediction task based on MIMIC-IV confirmed that client batch size serves as an effective mechanism for reducing privacy leakage while maintaining predictive utility. At batch size one, both one-hot architectures yielded near-perfect reconstruction at model initialization, retaining final attack point accuracies of 0.669 for ResNet and 0.850 for the small multilayer perceptron. As batch sizes grew to 128 and 256, reconstruction dropped substantially across both models. Synchronized evaluations of predictive performance revealed that larger batches did not compromise utility: ResNet validation area under the receiver operating characteristic curve increased from 0.830 at batch size one to 0.863 at batch size 32, while reconstruction fell from 0.669 to 0.504. The small multilayer perceptron displayed an identical trajectory, boosting utility from 0.823 to 0.863 while cutting reconstruction from 0.850 to 0.480. Exposure trajectories showed that although larger batch configurations began with lower predictive utility during early passes, they steadily improved over client data exposure, achieving competitive final utility alongside sustained reductions in gradient leakage.

Evaluating the FT-Transformer architecture revealed markedly different vulnerability characteristics compared to one-hot baselines. Across both benchmark and clinical datasets, FT-Transformer exhibited consistently lower reconstruction accuracy, but its leakage trajectory was non-monotonic across batch sizes. On MIMIC-IV, FT-Transformer reconstruction peaked at an intermediate batch size of 16, recording 0.426 at initialization and 0.462 at the final attack milestone, before tapering at larger batch sizes. Furthermore, complete record recovery diverged dramatically from aggregate feature recovery. At batch size one on MIMIC-IV, the final exact match rate reached 0.500 for ResNet and 0.700 for the small multilayer perceptron, whereas FT-Transformer achieved an exact match rate of 0.000. Because MIMIC-IV contains extensive sparse clinical history indicators and repetitive marginal profiles, raw feature accuracy was substantially elevated by recovering prevalent baseline values, where client marginal priors reached approximately 0.94. The findings demonstrated that aggregate reconstruction metrics can create an illusion of severe row-level compromise in sparse tabular settings, underscoring the absolute necessity of evaluating exact match rates and baseline offsets.

Extending the evaluation to FedAvg showed that multi-step local training trajectories do not function as a dependable, uniform privacy defense. When clients were restricted to 16 local examples, initialized FedAvg updates remained highly vulnerable, with ResNet reconstruction spanning 0.683 to 0.894 and the small multilayer perceptron spanning 0.738 to 0.812, while FT-Transformer remained between 0.262 and 0.335. Counterintuitively, increasing local computation from one to four epochs raised initialized reconstruction across all three architectures, proving that iterating repeatedly over small private record sets intensifies gradient signals rather than obscuring them. To determine whether FT-Transformer’s resilience stemmed solely from its categorical embedding interface, the researchers evaluated an embedding multilayer perceptron control matching FT-Transformer’s input representation. Across eight matched experimental settings on Adult and MIMIC-IV, the embedding multilayer perceptron proved more reconstructable than FT-Transformer in all eight instances, and exceeded the one-hot multilayer perceptron in six. These findings confirmed that learned categorical embeddings alone cannot explain the lower inversion success observed in transformer architectures.

Beyond comparisons across model families, an exhaustive hyperparameter grid within the multilayer perceptron architecture demonstrated that internal structural choices directly alter gradient leakage under identical federated training protocols. Analyzing variations across hidden widths of 32, 64, and 128, depths of one to three layers, LayerNorm versus BatchNorm, ReLU versus GELU activations, and dropout revealed pronounced effects on MIMIC-IV. Increasing hidden width consistently increased vulnerability, elevating final reconstruction from 0.473 to 0.526 at batch size eight and from 0.334 to 0.409 at batch size 32. Activation functions served as the primary module-level driver: switching from GELU to ReLU reduced average reconstruction from 0.588 to 0.413 at batch size eight, and from 0.422 to 0.321 at batch size 32. LayerNorm and dropout of 0.1 provided additional reductions in vulnerability. Across the complete batch size eight architecture grid, trajectory-averaged reconstruction varied from 0.208 for a three-layer LayerNorm ReLU network with dropout to 0.799 for the most vulnerable configuration, demonstrating a 59.1 percentage point margin attributable entirely to standard neural network design decisions.

The researchers emphasized critical limitations governing the interpretation of their findings, cautioning against viewing reduced reconstruction as an inherent privacy guarantee. The investigation evaluated an honest-but-curious server executing a specific optimization-based attack from LeakPro, omitting active attacks involving malicious parameter manipulation, client collusion, or generative priors. Furthermore, the experiments analyzed noncryptographic federated communication where individual client updates are directly exposed before aggregation. Production environments implementing secure aggregation fundamentally alter adversary observability by hiding individual updates within multi-client sums, while formal differential privacy bounds leakage through noise addition and gradient clipping. Similarly, local optimizer adaptations, server momentum, and post-processing defenses were excluded to maintain controlled comparisons across architectures. Ultimately, the authors concluded that while larger batch sizes, ReLU activations, normalization choices, and transformer architectures meaningfully suppress empirical reconstruction, they do not provide formal privacy protection.

Subject of Research: Privacy preservation and gradient inversion attacks in tabular federated learning

Article Title: Profiling privacy preservation against gradient inversion attacks in tabular federated learning

Article References: Nilsson, I. Ö., Engvall, M. B., Valadi, V., & Lazebnik, T. (2026). Profiling privacy preservation against gradient inversion attacks in tabular federated learning. International Journal of Information Security, 25(6), Article 192. https://doi.org/10.1007/s10207-026-01360-9

Image Credits: AI Generated

DOI: 10.1007/s10207-026-01360-9

Keywords: federated learning, gradient inversion attacks, tabular data, data privacy, honest-but-curious server, FedSGD, FedAvg, FT-Transformer, multilayer perceptron, ResNet, MIMIC-IV, exact match rate, batch size, model deltas, categorical embeddings, leakage profiling, machine learning security

Cite Scienmag News

Veronica Carney. (November 21, 2026). Tabular Federated Learning Updates Expose Private Records Under Gradient Inversion Attacks. Scienmag. https://scienmag.com/tabular-federated-learning-updates-expose-private-records-under-gradient-inversion-attacks/

Veronica Carney. "Tabular Federated Learning Updates Expose Private Records Under Gradient Inversion Attacks." Scienmag, 21 November 2026, https://scienmag.com/tabular-federated-learning-updates-expose-private-records-under-gradient-inversion-attacks/. Accessed 21 November 2026.

Veronica Carney. "Tabular Federated Learning Updates Expose Private Records Under Gradient Inversion Attacks." Scienmag. November 21, 2026. https://scienmag.com/tabular-federated-learning-updates-expose-private-records-under-gradient-inversion-attacks/

Tags: batch sizecategorical embeddingsData Privacyexact match rateFedAvgfederated learningFedSGDFT-Transformergradient inversion attackshonest-but-curious serverleakage profilingmachine learning securityMIMIC-IVmodel deltasmultilayer perceptronResNettabular data
Share26Tweet16
Previous Post

Deep Learning Covid-19 CT Models Show High Internal Accuracy but Generalization Limits

Next Post

Folate-Modified Albumin Nanoparticles Deliver Curcumin to Trigger Glioma Cell Death

Related Posts

Blue Colloidal Quantum Dots Advance Toward Commercial Light Emitting Displays
Science News

Blue Colloidal Quantum Dots Advance Toward Commercial Light Emitting Displays

November 21, 2026
Genetic Modifiers Alter Breast Cancer Risk in BRCA1 and BRCA2 Mutation Carriers
Cancer

Genetic Modifiers Alter Breast Cancer Risk in BRCA1 and BRCA2 Mutation Carriers

November 21, 2026
Crystal Defect Models Link Nanoscale Obstacles to Collective Dislocation Dynamics in Alloys
Technology and Engineering

Crystal Defect Models Link Nanoscale Obstacles to Collective Dislocation Dynamics in Alloys

November 21, 2026
Study Reveals How Loss of Heme Oxygenase Impairs Cellular Inflammatory Signaling
Science News

Study Reveals How Loss of Heme Oxygenase Impairs Cellular Inflammatory Signaling

November 21, 2026
Study Quantifies How Deep Rock Fracture Permeability Declines Under Normal Stress
Science News

Study Quantifies How Deep Rock Fracture Permeability Declines Under Normal Stress

November 21, 2026
Claudin 18.2 CAR T Therapy Shows Peritoneal Control in Phase I Gastric Cancer Cases
Cancer

Claudin 18.2 CAR T Therapy Shows Peritoneal Control in Phase I Gastric Cancer Cases

November 21, 2026
Next Post
Folate-Modified Albumin Nanoparticles Deliver Curcumin to Trigger Glioma Cell Death

Folate-Modified Albumin Nanoparticles Deliver Curcumin to Trigger Glioma Cell Death

  • Mothers who receive childcare support from maternal grandparents show more optimized

    Mothers who receive childcare support from maternal grandparents show more parental warmth, finds NTU Singapore study

    27656 shares
    Share 11059 Tweet 6912
  • University of Seville Breaks 120-Year-Old Mystery, Revises a Key Einstein Concept

    1061 shares
    Share 424 Tweet 265
  • Bee body mass, pathogens and local climate influence heat tolerance

    682 shares
    Share 273 Tweet 171
  • Researchers record first-ever images and data of a shark experiencing a boat strike

    546 shares
    Share 218 Tweet 137
  • Groundbreaking Clinical Trial Reveals Lubiprostone Enhances Kidney Function

    531 shares
    Share 212 Tweet 133
Science

Embark on a thrilling journey of discovery with Scienmag.com—your ultimate source for cutting-edge breakthroughs. Immerse yourself in a world where curiosity knows no limits and tomorrow’s possibilities become today’s reality!

RECENT NEWS

  • Blue Colloidal Quantum Dots Advance Toward Commercial Light Emitting Displays
  • Genetic Modifiers Alter Breast Cancer Risk in BRCA1 and BRCA2 Mutation Carriers
  • Crystal Defect Models Link Nanoscale Obstacles to Collective Dislocation Dynamics in Alloys
  • Study Reveals How Loss of Heme Oxygenase Impairs Cellular Inflammatory Signaling

Categories

  • Agriculture
  • Anthropology
  • Archaeology
  • Athmospheric
  • Biology
  • Biotechnology
  • Blog
  • Bussines
  • Cancer
  • Chemistry
  • Climate
  • Earth Science
  • Editorial Policy
  • Marine
  • Mathematics
  • Medicine
  • Pediatry
  • Policy
  • Psychology & Psychiatry
  • Science Education
  • Science News
  • Social Science
  • Space
  • Technology and Engineering

Subscribe to Blog via Email

Enter your email address to subscribe to this blog and receive notifications of new posts by email.

Join 5,151 other subscribers

© 2026 Scienmag - Science Magazine

Welcome Back!

Login to your account below

Forgotten Password?

Retrieve your password

Please enter your username or email address to reset your password.

Log In
No Result
View All Result
  • HOME
  • SCIENCE NEWS
  • CONTACT US

© 2026 Scienmag - Science Magazine

Discover more from Science

Subscribe now to keep reading and get access to the full archive.

Continue reading