Research map / Federated learning and privacy / Privacy attacks and secure learning
FACT or Fiction: Can Truthful Mechanisms Eliminate Federated Free Riding?
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 FACT, a federated learning mechanism that uses a free-rider penalty plus a random "sandwich" truthfulness competition so agents report true costs and use locally optimal data, evaluated via simulated and real federated experiments on CIFAR-10, MNIST, and HAM10000.
Problem
Standard federated learning suffers from free riding, and existing anti-free-riding mechanisms assume agents truthfully report their costs; untruthful agents can misreport costs to free ride.
Approach
A free-rider penalty shifts the federated optimum back to each agent's local optimum, and a randomized sandwich competition (agents win if their reported cost lies between two randomly sampled costs, with rewards paid from other agents' contract fees) makes reporting the true cost each agent's best strategy, without auctions or bandits.
Key results
- FACT reduces agent loss compared to local training by up to 3x on CIFAR-10 and 4x on MNIST.
- Agents' net improvement in loss peaks when they report their true cost (0% inflation) and decreases for inflated or deflated reports, in iid and non-iid settings.
- On the HAM10000 skin-lesion case study, participating agent loss is reduced by nearly 66% compared to local training.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FACT instead of local training | 16-agent federated image classification · CIFAR10 | agent loss reduction vs local training: up to a factor of 3x | Helped |
| Use FACT instead of local training | 16-agent federated image classification · MNIST | agent loss reduction vs local training: 4x | Helped |
| Use FACT instead of local training | ResNet50 pretrained on ImageNet · HAM10000 | agent loss reduction vs local training: nearly 66% | Helped |
| Add the free-riding penalty that targets locally optimal data contribution | FACT · CIFAR10 and MNIST | free-riding at the data-contribution optimum | Helped |
| Use the sandwich truthfulness competition | FACT · CIFAR10 and MNIST, iid and non-iid | net improvement in loss when reporting costs | Helped |
Limitations
- Assumes the central server is honest and agents are rational, non-colluding, and not malicious in gradient updates.
- Theory restricted to iid data; non-iid mechanism left to future work.
- Truthfulness competition is evaluated via simulated sampled costs rather than real agent behavior; no baselines beyond local training.
Experimental setup
Datasets: CIFAR-10 MNIST HAM10000
Benchmarks: CIFAR-10 MNIST HAM10000
Metrics: agent loss net improvement in loss over local training test loss test accuracy
Baselines: local training traditional FedAvg federated learning
Builds on: FedAvg Dirichlet non-iid partitioning
Closest papers by meaning
- Incentivized Truthful Communication for Federated Bandits (ICLR 2024 · similarity 0.86)
- You Get What You Give: Reciprocally Fair Federated Learning (ICML 2025 · similarity 0.84)
- Stochastically Dominant Peer Prediction (NeurIPS 2025 · similarity 0.83)
- NoT: Federated Unlearning via Weight Negation (CVPR 2025 · similarity 0.82)
- A Cramér–von Mises Approach to Incentivizing Truthful Data Sharing (NeurIPS 2025 · similarity 0.82)
- Towards Fair Graph Federated Learning via Incentive Mechanisms (AAAI 2024 · similarity 0.82)
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.