# Does the Trust Boundary Hold over Time? A Long-Horizon Benchmark for Receiver-Directed Federated Learning

DevFeed: [Does the Trust Boundary Hold over Time? A Long-Horizon Benchmark for Receiver-Directed Federated Learning](<https://devfeed.tech/articles/does-the-trust-boundary-hold-over-time-a-long-horizon-benchmark-for-receiver-directed-federated-learning-116477.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.03724>)

Author: Ting Xu, Henry Leung

Published: 2026-11-04T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Distributed Computing](<https://devfeed.tech/sources/arxiv-distributed-computing.md>), [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [Federated Learning](<https://devfeed.tech/topics/federated-learning.md>), [Machine learning](<https://devfeed.tech/topics/machine-learning.md>), [Adversarial attacks](<https://devfeed.tech/topics/adversarial-attacks.md>), [audit](<https://devfeed.tech/topics/audit.md>)

Tags: [adversarial](<https://devfeed.tech/tags/adversarial.md>), [audit](<https://devfeed.tech/tags/audit.md>), [benchmark](<https://devfeed.tech/tags/benchmark.md>), [cpu](<https://devfeed.tech/tags/cpu.md>), [cs-dc](<https://devfeed.tech/tags/cs-dc.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [federated-learning](<https://devfeed.tech/tags/federated-learning.md>), [gate](<https://devfeed.tech/tags/gate.md>), [horizon](<https://devfeed.tech/tags/horizon.md>), [model](<https://devfeed.tech/tags/model.md>), [models](<https://devfeed.tech/tags/models.md>), [oracle](<https://devfeed.tech/tags/oracle.md>), [over](<https://devfeed.tech/tags/over.md>), [protocol](<https://devfeed.tech/tags/protocol.md>), [qwen3](<https://devfeed.tech/tags/qwen3.md>), [receiver](<https://devfeed.tech/tags/receiver.md>), [sites](<https://devfeed.tech/tags/sites.md>), [synthetic](<https://devfeed.tech/tags/synthetic.md>), [target](<https://devfeed.tech/tags/target.md>), [training](<https://devfeed.tech/tags/training.md>), [trust](<https://devfeed.tech/tags/trust.md>), [useful](<https://devfeed.tech/tags/useful.md>)

## AI overview

The benchmark finds that a receiver's trust boundary can fail over long training, depending on its reference point. Across 20 paired CPU seeds, judging transfers against a fixed starting model better identified starting-model improvements but admitted 29% harmful to the receiver's current model. Judging against the current model admitted 2% harmful transfers on that target, while missing many gains relative to the starting model. A held-out audit leaves gate verdicts unchanged and, after training, chooses deployment or fallback to local training. At 600 rounds it deployed 65-72% of models, kept 98% of the gain, and none had negative benefit on the audit target; with 20,000 audit observations, it detected 99% of stale transfers costing at least 0.01 log loss. In the reported attack setting with 38% adversarial sites, FedFomo and CS-Gate were the only evaluated methods with positive mean gain in every honest group.

## Source excerpt

arXiv:2611.03724v1 Announce Type: cross Abstract: In receiver-directed federated learning, each site draws a trust boundary. For every source and model component it judges, on its own data, whether the transfer helps (admit), harms (reject) or cannot yet be told (abstain). We ask whether this boundary holds over hundreds of rounds. Our benchmark has a federation of thirteen semi-synthetic sites, an exact oracle for every evaluated transfer, paired seeds, an attack protocol, fourteen methods over 600 rounds and Qwen3-1.7B LoRA sites over 500. The answer depends on which boundary is measured. Across 20 paired CPU seeds, a gate judging against a fixed reference better identified transfers useful relative to the site's starting model. Yet it admitted 29% of transfers harmful to the model the site now has. A gate judging against the current model admitted 2% of harmful transfers on that target in the CPU tier, but missed many transfers useful to the starting model. Two short propositions explain these distinct failures. We also evaluate a held out audit that leaves gate verdicts intact. After training, it decides whether each final model is deployed or the site falls back to local training, and reports admissions that have gone stale. With independent audit data its guarantees hold after any training history. At 600 rounds the audit deployed 65 to 72% of models, none with negative benefit on its target, and kept 98% of the gain. No stale report had the wrong sign in our runs. With 20,000 audit observations, a variance-adaptive bound found 99% of stale transfers that cost at least 0.01 in log loss. A common-target Qwen3 comparison on 60 selected transfers shows the same tension over 100 rounds. With 38% adversarial sites, FedFomo and CS-Gate were the only evaluated methods with positive mean gain in every honest group.