Research map / Federated learning and privacy / Communication-efficient FL
DICE: Data Influence Cascade in Decentralized Learning
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 DICE, the first method to estimate how a data instance's influence cascades across multi-hop neighbors in decentralized learning, via gradient-based approximations linking influence to topology, data, and loss curvature.
Problem
Quantifying individual data contributions (influence) in fully decentralized learning, where influence propagates through peer-to-peer gossip to multi-hop neighbors, which static centralized influence estimators cannot capture.
Approach
Defines a ground-truth multi-hop influence cascade (DICE-GT) based on leave-one-out-style removal across neighbors, and derives tractable first-order gradient-based approximations (DICE-E) showing influence depends on communication weights, neighbor curvature (Hessians), and the local update; applies proximal influence and reciprocity factors for collaborator selection and anomaly detection.
Key results
- One-hop DICE-E closely tracks ground-truth DICE-GT across CIFAR-10, CIFAR-100 and Tiny ImageNet with CNN and MLP models on a 32-node ring graph (30 points per plot).
- In anomaly detection on a 32-node exponential graph with CIFAR-10, malicious label-flipping neighbors are readily detectable via high proximal influence.
- Influence of an identical data batch is visibly larger on high out-weight (topologically important) nodes, showing topology-dependent asymmetry.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use one-hop DICE-E as an approximation of one-hop DICE-GT | DICE on CIFAR-10, CIFAR-100, and Tiny ImageNet with CNN and MLP models · CIFAR-10, CIFAR-100, Tiny ImageNet | influence alignment | Helped |
| Use proximal influence to identify neighbors with harmful contributions | DICE anomaly detection on an exponential graph · CIFAR-10 | anomaly detection | Helped |
| Place the same data batch at a node with greater topological importance | One-hop DICE-E on decentralized ResNet-18 training · CIFAR-10 | influence score | Helped |
Surprising findings
- Influence depends on topological importance and neighbor curvature, not just the data and local model, and decays exponentially with hop count.
Limitations
- Multi-hop approximation is only derived in theory and validated empirically only for one-hop; higher-order experiments are not shown.
- Approximations rely on first-order Taylor expansion and small learning rates.
- Experiments are small-scale image classification on 32-node graphs with 512-sample subsets.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 Tiny ImageNet
Benchmarks: CIFAR-10 CIFAR-100 Tiny ImageNet
Metrics: alignment between DICE-GT and DICE-E
Baselines: TracIn
Builds on: TracIn leave-one-out influence decentralized SGD In-Run Data Shapley
Closest papers by meaning
- Competitive Advantage Attacks to Decentralized Federated Learning (NeurIPS 2025 · similarity 0.84)
- Peer Neighborhood Mechanisms: A Framework for Mechanism Generalization (AAAI 2024 · similarity 0.84)
- Privacy Attacks in Decentralized Learning (ICML 2024 · similarity 0.84)
- Statistical Collusion by Collectives on Learning Platforms (ICML 2025 · similarity 0.83)
- Decentralized Ranking Aggregation via Gossip: Convergence and Robustness (NeurIPS 2026 · similarity 0.83)
- You Get What You Give: Reciprocally Fair Federated Learning (ICML 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.