Research map / Federated learning and privacy / Personalized and heterogeneous FL

HaCore: Efficient Coreset Construction with Locality Sensitive Hashing for Vertical Federated Learning

AAAI 2025 · Federated learning and privacy › Personalized and heterogeneous FL · approach: Vertical federated learning · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes HaCore, which builds VFL coresets by clustering locality-sensitive-hashing binary signatures with k-medoids accelerated by an inverted index and bit-counting.

Problem

Vertical federated learning training is costly with many samples; existing VFL coreset methods (V-coreset, TreeCSS) produce overly large coresets with many clients, have long construction times, and lack generality across models.

Approach

Map features to LSH bit signatures locally on each client, concatenate them into global signatures, and run Hamming-distance k-medoids so medoid samples form the coreset; an inverted index on medoids and a bit-counting method accelerate assignment and medoid selection.

Key results

• HaCore reaches 76.71% on SUSY with LR vs 76.93% for full-data training at 2% coreset size.
• HaCore accelerates the best baseline TreeCSS by up to 45x (on HIGGS) and by about 28x on average on other datasets in end-to-end time.
• Bit-counting and inverted-index optimizations reduce end-to-end time on SUSY from 3118s to 772s with essentially unchanged accuracy.

Measured effects reported in the paper

Surprising findings

• Using a smaller coreset degrades accuracy for TreeCSS, and TreeCSS becomes unstable as the number of clients changes.

Limitations

• Experiments use only four clients and simple MLP/LR models, as CNNs are rarely used in VFL.
• Privacy analysis is under an honest-but-curious assumption and does not address server-client collusion; stricter privacy requires homomorphic encryption at efficiency cost.

Experimental setup

Datasets: SUSY HIGGS Covertype YearPredictionMSD House Prices

Benchmarks: SUSY HIGGS Covertype YearPredictionMSD House Prices

Metrics: accuracy RMSE end-to-end training time

Baselines: Full Random TreeCSS V-coreset

Builds on: SplitNN locality sensitive hashing k-medoids TreeCSS V-coreset

Closest papers by meaning

• Improved Coresets for Vertical Federated Learning: Regularized Linear and Logistic Regressions (ICML 2025 · similarity 0.88)
• VFLAIR: A Research Library and Benchmark for Vertical Federated Learning (ICLR 2024 · similarity 0.85)
• Low Precision Local Training is Enough for Federated Learning (NeurIPS 2024 · similarity 0.84)
• FedCS: Coreset Selection for Federated Learning (CVPR 2025 · similarity 0.83)
• Vertical Federated Learning with Missing Features During Training and Inference (ICLR 2025 · similarity 0.83)
• FedWon: Triumphing Multi-domain Federated Learning Without Normalization (ICLR 2024 · similarity 0.83)