Telling Poisoned Updates Apart from Honest Domain Shift in Federated Information Extraction

Suppose several organizations want to jointly train a model that reads text and pulls out entities and the relations between them, for example who works for whom or which drug causes which side effect. None of them will share raw text, each writes in a different genre, and some participants may be dishonest. Two existing tools address parts of this problem, and they fail in opposite ways. Reputation systems in decentralized learning judge an update by how well it performs on validators' own data. Under domain shift, that test cannot separate a malicious update from an honest one trained on unusual text, so it tends to expel exactly the participants whose data is most worth having. Domain-adaptation weighting does the reverse. It prefers updates whose features look compatible with the target genre and has no notion of honesty, so a poisoner whose features cluster well is accepted.

We propose a shift-corrected reputation score that sits between these two failures. A validation committee scores each update on hidden test sentences, much as an oracle network is audited with secret queries. The committee does not use that score directly. It compares the score against the score an honest participant at the same measured domain distance would be expected to get, and only the unexplained shortfall lowers reputation. The domain distance comes from the same domain discriminator that already drives adversarial adaptation. The same distance also lowers the update's aggregation weight. A cheater who pretends to come from a distant domain to excuse a low score therefore buys tolerance only by giving up influence.

Core Claim

The failure we target is specific and has not been tested. In cross-genre federated joint extraction, validation-based reputation of the kind used by BFLC will flag honest distant-genre participants as faulty, and this will hurt target-genre accuracy even when no attacker is present. Two simple changes fix it without giving up robustness to poisoning. The first is subtracting the score expected at a given domain distance. The second is scoring entity and relation outputs separately. We expect the corrected score to match per-domain oracle committees on clean data and to match BFLC-style filtering under attack.

Positioning Against Prior Work

The proposal reuses three methods. The first is the gradient reversal layer: a layer that passes features forward unchanged but flips the sign of the gradient coming back from a domain classifier, which pushes the encoder toward genre-invariant features. Fu et al. applied it to relation classification across ACE 2005 genres. Their model assumed gold entity spans and kept all data in one place, and they found that aligning all target genres at once performed worse than aligning each separately. We treat that negative result as a constraint: global alignment erases genre-specific signal. For this reason we keep a per-participant domain distance rather than trying to drive it to zero.

The second reused method is a joint extraction model with contextualized span representations, in the style of DYGIE++. We also run a pipeline variant built on PURE, because Zhong and Chen showed that separate entity and relation encoders beat shared ones. Their result directly challenges any design that puts one adversarial discriminator on one shared encoder. We respond by testing both backbones, and by placing separate discriminators on the entity and relation representations in the joint model.

The third reused method is an asymmetric reputation update from reputation-based state machine replication. Their bounded, tanh-shaped update makes reputation slow to gain and fast to lose. We use it as is, together with their observation that honest nodes must hold consistent reputation views.

Two lines of work motivate the gap from opposite sides. Federated Adversarial Domain Adaptation (FADA) weights source participants by how cleanly their features cluster ("gap statistics"). That measures compatibility, not honesty. BFLC filters updates through committee validation scores and stays above 80% accuracy on FEMNIST with 30% malicious clients, while plain federated learning falls below 20%. However, every BFLC participant shares one image domain, so the confusion between shift and malice that we study never arises there.

DecTest supplies the idea of covert test queries for auditing oracles, which we adapt into hidden test sentences with known gold triples. SC-PBFT and BLOCKS combine several signals into a trust score: latency and throughput in the first, downstream utility for retrieval in the second. Neither corrects for distribution shift, and neither trains model representations. CrossRE shows how large genre variance is: Micro-F1 is 75.96 on Music but 40.10 on Natural Science, and average Macro-F1 across its six domains is 28.29. A gap that wide is precisely what makes raw validation scores misleading.

We do not claim a new consensus protocol. The ledger and the committee election follow BFLC. Our contribution is the scoring rule and the evidence about when it is needed.

Mechanism

In each round, participant $i$ submits an update. The committee runs it on a hidden test set drawn from the genres the committee holds. From this it records an entity score $s_i^{E}$ and a relation score $s_i^{R}$, both strict F1. It also records a domain distance $d_i$, defined as the held-out accuracy of the global domain discriminator at telling the participant's encoded unlabeled sample apart from the pooled reference pool.

Across all participants in the round, the committee fits a robust monotone curve $g^{k}$ for each head $k \in \{E, R\}$. The curve maps distance to expected score and is fit by monotone regression with a median loss, so that a minority of attackers cannot bend it. The shortfall for each head is then

$$r_i^{k} = \min\big(0,\; s_i^{k} - g^{k}(d_i)\big).$$

Reputation $\mu_i$ is updated with the tanh rule using the worse of the two shortfalls. The worse one is used because a targeted attack on relation labels can leave entity scores untouched. The aggregation weight is

$$w_i = \mu_i \cdot \exp(-\beta d_i),$$

where $\beta$ is a temperature tuned on clean development rounds. This weight is the trade-off described above: claiming a large $d_i$ shrinks the shortfall but also shrinks $w_i$.

Research Program

First, we build a simulator for federated joint extraction. Each ACE 2005 genre (newswire, broadcast news, broadcast conversation, telephone conversation, weblogs, usenet) is split across several participants with labeled data. One held-out genre serves as the target, with unlabeled text only. CrossRE's six domains form a second testbed, with joint extraction over its entity and relation annotations.

Second, we measure the core premise directly. On clean data, we check how often BFLC-style raw scoring drops honest distant-genre participants and what that exclusion costs on the target genre.

Third, we evaluate four attacks. The first flips the labels of one relation type. The second plants a backdoor in which a trigger entity name forces a chosen relation, echoing the spurious "Obama"-to-employment correlation in Fu et al.. The third submits random-noise updates. The fourth is an adaptive "camouflaged" attacker that makes its features look genre-distant to excuse a low score.

Fourth, we run the PURE-style pipeline backbone alongside the joint backbone, to see whether the correction depends on whether representations are shared.

Baselines

The naive reference is federated averaging with a gradient reversal layer on the shared encoder and no filtering. It shows what alignment alone does under attack.

The strongest honesty-based competitor is BFLC committee scoring with median thresholding, run on the identical model, data partition, number of rounds, and committee size. BFLC reports above 80% accuracy with 30% malicious clients on FEMNIST. There is no reported result on Lacuna for BFLC on any text extraction task.

The strongest compatibility-based competitor is FADA's gap-statistic attention weighting combined with the same reversal layer. We reimplement it for span features, since its reported results are on images and sentiment reviews (73.6% on Digit-Five, for example).

For the classical robust statistics option, we use coordinate-wise median and trimmed mean aggregation. The file notes that BFLC compared against coordinate-wise median, but gives no number for it, so there is no reported result found on Lacuna for coordinate-wise median on FEMNIST or on extraction.

For absolute accuracy, the closest anchors are centralized ones. Fu et al.'s adversarially adapted relation classifier, which uses gold entities, reaches 65.16 F1 on broadcast conversation, 55.55 on weblogs, and 57.19 on telephone conversation, starting from source-only CNN scores of 64.44, 54.58, and 57.02. Bekoulis et al.'s joint model reaches 64.54 F1 on ACE04. There is no reported result found on Lacuna for federated joint extraction on ACE 2005 genres. We therefore also run a centralized model with reversal-layer adaptation, trained on all clean data, as an upper bound.

All methods share the same encoder, the same tuning budget (an equal number of configurations, chosen on a clean development target genre that is never the test genre), and the same communication rounds.

Two ablations isolate the key ingredient. The first removes the distance correction by setting $g$ to a constant, which turns the rule back into BFLC with separate heads. The second merges entity and relation scores into one.

We count the proposal as winning if two conditions hold. On clean data, it must keep at least as many honest distant-genre participants as uniform averaging and reach target strict relation F1 within one point of the clean centralized upper bound. Under 30% attackers, it must stay within one point of BFLC on attack success rate while beating BFLC on target F1.

Why not just…?

Why not just use coordinate-wise median or trimmed mean? These rules discard updates that sit far from the bulk on each coordinate. When genres are unevenly represented, an honest minority genre is exactly such an outlier. We therefore expect median-style rules to fail in the same way raw validation does, and to be blind to backdoors that stay within normal coordinate ranges. The clean-data honest-retention measurement settles this. If the median retains minority genres and holds target F1, our correction is unnecessary.

Why not just use FADA's gap-statistic weighting? Gap statistics reward feature compactness, and a relation label-flip attack leaves encoder features largely intact. The label-flip and backdoor attacks test this directly. FADA should match us on clean data and lose under attack. If it does not lose, the honesty half of the design adds nothing.

Why not just build committees from the same genre as the participant being judged? This is the most serious threat. When every genre has several labeled participants, same-genre validation removes the confusion between shift and malice entirely. We include it as an oracle baseline. Our argument for the correction rests on regimes where this oracle is unavailable: a genre held by a single participant, or a genre the committee has no labels for. We will report results with the number of participants per genre varied from one to several, and say plainly if the oracle dominates once two or more exist.

Why not just skip adaptation and use domain-adaptive pretraining? Continued language-model pretraining on target text can raise accuracy, as SciBERT does for DYGIE++. It does nothing about filtering, though, and it does not remove spurious entity-to-relation correlations of the kind Fu et al. report. We run it as an additional backbone under every aggregation rule. If its gains make the reversal layer redundant, the domain distance can come from a separately trained probe instead. The scoring rule does not need the reversal layer to be the source of $d_i$.

Decisive Checks

The first check tests the premise. With no attackers, on ACE 2005 with one target genre held out, BFLC-style raw scoring should exclude honest telephone-conversation or weblog participants at clearly higher rates than newswire participants, and its target F1 should fall below uniform averaging. If neither happens, the problem we propose to solve does not exist at this scale, and we will report that as the finding.

The second check is the camouflage test. Against the adaptive attacker, the attacker's effective influence, measured as the sum of its aggregation weights, should fall as it inflates its apparent domain distance. Its attack success rate should not exceed the rate it achieves against BFLC.

The third check is per-head scoring. On the relation label-flip attack, the merged-score ablation should miss attackers whose entity F1 stays normal, and the separate-head rule should catch them.

All three are evaluated on both ACE 2005 and CrossRE, with strict entity-boundary-plus-type-plus-relation F1 as the primary metric.