Federated Learning Model Governance: Who Owns the Weights, Updates, and Errors

The Coopetition Problem: Why Federated Learning Broke Traditional Pharma Contracts

Pharma has always known how to write contracts for collaboration. Joint ventures, co-development agreements, in-licensing deals, and public-private partnerships are old craft. What is genuinely new about federated learning is that it turns competitors into shared owners of an ongoing, mutable asset without ever moving the underlying data that gave the asset its value.

In a conventional research consortium, either the data sits together in a shared warehouse (and legal teams negotiate access, purpose limitation, and audit rights over that warehouse), or the parties publish separate contributions and split the credit. Federated learning collapses both patterns. The data never moves, so classical data-sharing agreements over-protect what does not need protecting. And the “contribution” each party makes is a stream of encrypted gradient updates that materially changes the shared model but is invisible to counterparties. There is no single artifact to hand back at termination, no clean line between what belonged to whom, and no obvious way to reason about who is responsible for a shared output that keeps evolving.³

MELLODDY, the ten-pharma consortium that trained a shared drug-discovery model across proprietary chemistry libraries, is the closest thing the industry has to a legal precedent. The consortium’s own writeups describe it as “turning competition into coopetition,” and the technical partners published extensively on the platform. What they published less about was the twelve months of contracting that came before any model was trained: legal terms had to cover IP ownership over derived models, gradient exchange as pseudonymised data under GDPR, and downstream commercialisation rights.⁴¹

The governance question is not whether federated learning is legally possible. It clearly is. The question is what the participant agreements, dispute mechanisms, and regulatory allocations should look like now that the pattern is repeating across cheminformatics, imaging, real-world evidence, and post-market surveillance. A one-off legal team can write a bespoke agreement for one consortium. The industry needs reusable templates that survive the next ten.

Who Owns the Weights? Untangling Global, Local, and Contribution-Weighted IP

Ownership is the first question executives ask and often the last one legal answers. In a federated learning consortium, at least four distinct artifacts exist, and each raises a different ownership problem.

First, the local model each participant trains inside its own firewall on its own data. This is almost always the participant’s sole property, and every serious FL agreement should say so explicitly, including a covenant that the coordinator will never receive local model weights, only gradient contributions.

Second, the gradient updates exchanged during each training round. Legally, these are usually treated as pseudonymised data derived from the local training set, which means they carry the same protection level as regular data sharing and require a lawful basis under GDPR or equivalent frameworks.⁶ They are ephemeral, but they are also the substrate on which the shared IP is built.

Third, the global model that emerges from aggregation. This is the asset that fuels most disputes. Options for ownership include joint ownership by all participants in equal shares, joint ownership weighted by contribution, sole ownership by a consortium legal entity that licenses it back to members, or a hybrid where the model is owned by the consortium but usage rights are granted to participants under specified fields-of-use.⁴

Fourth, the derived works: fine-tuned variants each participant creates from the global model for its own internal use. Whether these carry a residual license fee, revert to the consortium at termination, or belong outright to the fine-tuner is one of the most negotiated terms in FL contracting.

Watermarking and ownership verification schemes such as FedIPR, FedRight, and FedTracker are increasingly proposed as technical enforcement layers on top of the legal ownership scheme. FedTracker, for example, provides both ownership verification and traceability for federated models by embedding participant-specific fingerprints that survive fine-tuning.⁷ A well-drafted consortium agreement should reserve the coordinator’s right to embed such fingerprints and require participants not to remove them.

Contribution Attribution When Training Data Is Asymmetric

Every FL consortium eventually confronts the same uncomfortable truth: not every participant contributes equally. Some bring millions of high-quality records; others bring a few thousand noisy ones. Some bring rare disease cohorts that lift the global model far beyond its baseline; others bring redundant common-population data. Yet the naive federated-averaging algorithm treats all participants as equal contributors to each round.

Fair attribution matters for three reasons. It affects any revenue-share or downstream royalty allocation. It determines each participant’s leverage in governance votes. And when regulators or plaintiffs ask “whose data shaped this decision,” it is the only defensible answer.

The academic literature has converged on Shapley values as the fair-contribution metric of choice. Derived from cooperative game theory, Shapley values quantify each participant’s marginal contribution to the coalition’s total performance by evaluating every possible subset of participants and measuring how much each one adds. Papers such as GTG-Shapley, and vertical-federated-learning contribution schemes, have made Shapley calculation tractable at pharma scale, though costs remain high.⁸⁹

Whichever method is chosen, three governance rules should be codified in the participant agreement. Contribution must be calculated by an independent party, typically the technical operator, using an algorithm named in the agreement. It must be auditable, meaning any participant can demand an independent recalculation. And it must be subject to a minimum quality floor, so a participant flooding the consortium with low-quality data cannot game the attribution formula. Research has explicitly warned that Shapley-based contribution metrics can be volatile in non-IID settings and can be manipulated by adversarial participants.¹⁰

Contract Structures for FL Consortia: Four Working Models

The right legal structure depends on how commercial the consortium is, how many participants it has, and whether a third-party technical operator sits between the participants. Four patterns have emerged in the pharma federated-learning landscape.

Regardless of structure, every FL consortium agreement should include the same set of core schedules: (1) technical protocol and infrastructure specification, (2) IP and license schedule for local models, gradients, global model, and derived works, (3) contribution attribution methodology, (4) validation and update-approval process, (5) liability and indemnity allocation, (6) regulatory correspondence rules, (7) exit and termination consequences, and (8) dispute-resolution mechanism.

Liability When the Federated Model Makes a Wrong Decision

The most anxious question at every FL kickoff meeting is not who owns the model. It is who is liable when it is wrong. The answer is unsatisfying because it is genuinely novel legal territory.

Researchers have described the situation as a “double black box” problem. In centralized ML, opacity comes from the model itself; regulators and courts have learned to require validation documentation, performance monitoring, and explainability. In federated ML, a second layer of opacity is added: the coordinator cannot inspect the data of any single participant, so when the global model produces a spurious association or an outright error, it is often impossible to trace which participant’s local training set introduced the fault.¹²

Concrete example: a global FL model deployed for pharmacovigilance signal detection flags a false association between a marketed drug and a serious adverse event. Investigation reveals that one participant’s local data included a mis-coded ICD field for a period of six months. Under classical tort logic, that participant would bear responsibility. Under federated learning, the coordinator cannot see the participant’s data, the participant may deny fault, and by the time the audit happens, the offending data has been overwritten several rounds later. The trace is gone.

This structure requires two prerequisites. First, robust traceability: the coordinator must log every training round, every gradient contribution, and every model version, in a tamper-evident way (some consortia use blockchain-anchored hashes). Second, an agreed-upon forensic protocol: when a fault is alleged, an independent forensic auditor with contractual access to each participant’s local data (under NDA) attempts to trace the fault before the pro-rata fallback kicks in.

Insurance markets are beginning to respond. Cyber and product-liability carriers are quoting FL-specific endorsements that cover “algorithmic collaboration” scenarios, though rates are still highly variable and underwriters are wary of the unattributable-fault tier. Consortia should carry named-insured coverage at the entity level, with each participant carrying separate coverage for its own contributions.

Update Cycles: Who Approves, Who Tests, Who Deploys

A federated model is not shipped once and forgotten. It is continuously updated. In pharma settings, especially anything touching a regulatory submission, GxP-adjacent use, or clinical decision support, every update cycle needs governance.

The critical asymmetry to build into the agreement: a candidate release must be ratified by supermajority of the technical steering committee, but any single participant retains the unilateral right to withhold the release from its own environment. This “opt-out-per-cycle” clause is the safety valve that keeps regulated participants comfortable. It also aligns with the FDA and EMA’s January 2026 joint Guiding Principles on Good AI Practice, which emphasize risk-based lifecycle management and clear essential information across the product lifecycle.¹³

Handling Errors: Which Participant Fixes What

Errors in federated learning fall into three categories, and each category has a different owner.

Local data errors (mislabeled records, harmonization drift, upstream ETL bugs) are the responsibility of the contributing participant. The consortium’s role is to detect them: automated anomaly detection on gradient contributions, per-round performance monitoring by participant, and periodic on-site audit rights (typically once per year, with 30 days’ notice). When detected, the affected participant must remediate within an agreed window, and the coordinator must decide whether to retrain from a known-good checkpoint.

Aggregation errors (bugs in the FL framework, orchestration failures, secure-aggregation cryptographic issues) are the responsibility of the technical operator. Every FL agreement should carry a service-level agreement with the operator covering uptime, aggregation correctness, and vulnerability response.

Global model errors (poor generalization, subgroup underperformance, spurious associations) belong to the consortium collectively. Remediation is a governance decision: rollback, retrain, or accept-with-mitigations. The agreement should specify which body decides (usually the technical steering committee under a defined quorum), how quickly it must decide, and how participants are notified.

Regulatory Exposure Sharing Under FDA and EU AI Act

Regulatory exposure is where the boardroom sweat starts. A federated model used in support of a marketing authorization application, pharmacovigilance activity, or clinical decision support tool touches at least three regulatory frameworks simultaneously: the underlying medicinal-product framework (FDA CFR Title 21, EMA guidelines), the horizontal AI framework (FDA AI/ML guidance, EU AI Act), and the data-protection framework (HIPAA, GDPR, jurisdiction-specific health data laws).¹⁴¹⁵

The January 2026 joint FDA-EMA Guiding Principles of Good AI Practice in Drug Development lay out ten principles that map surprisingly well onto federated-learning governance, including risk-based approach, data governance and documentation, model design and development practices, risk-based performance assessment, and lifecycle management.¹³ None of the principles prohibit federated learning; all of them require that the sponsor using the model be able to demonstrate compliance for its specific context of use.

Under the EU AI Act, pharmaceutical AI systems used in diagnostics, patient monitoring, or clinical decision support are classified as high-risk under Annex III, with most obligations coming into force in August 2026 and a full compliance transition through August 2027 for high-risk AI in regulated products.⁵ Federated learning consortia therefore have a narrow window to build the technical documentation, risk management systems, and post-market monitoring processes the Act requires. EFPIA has argued that AI systems used solely for medicines R&D are exempt from the AI Act, but that argument does not extend to systems used in diagnostics, decision support, or manufacturing.¹⁶

Consortium agreements should therefore allocate regulatory exposure explicitly. Each participant is responsible for compliance in its own jurisdiction and use case; the consortium (or technical operator) is responsible for producing a common documentation package (technical file, data governance description, model card, validation results) that each participant can adapt to its specific filings.

FL Governance Charter Template and Participant Agreement Terms

The following is a practitioner’s charter template. It is not legal advice, but it distills what recurring FL consortia converge on in practice, based on public documentation from MELLODDY, the AISB Network, and Owkin’s hospital consortia.

FL Consortium Governance Charter (skeleton)

Participant agreement key terms

Each participant should sign a shorter participant agreement referencing the charter, plus schedules covering: data handling and infrastructure security, gradient contribution license, global model use license, indemnity, and dispute-resolution consent.

Dispute Resolution Framework for Federated Consortia

Disputes in FL consortia have a distinctive profile. Many are technical (contribution attribution recalculation, release-candidate performance disagreement, forensic-audit findings), some are commercial (fields-of-use disputes, indemnity claims, exit terms), and some are governance (voting deadlocks, expulsion). A one-size arbitration clause does not serve all three.

Recent academic work has begun to codify federated-learning-specific dispute resolution: a taxonomy of client-level disagreements in FL identifies scenarios such as strategic exclusion, regulatory carve-outs, and asymmetric competitive concerns that cannot be resolved by simply averaging gradients. Multi-track resolution strategies, including isolated model update paths, are proposed to accommodate these cases without breaking the federation.¹⁷ Consortia should reference these approaches in the dispute clause and empower the technical steering committee to implement isolated update paths as an interim remedy.

Conclusion

Federated learning does something genuinely useful for pharma: it lets competitors share the statistical power of their combined data without sharing the data. That capability is worth building consortia around. But the consortia only pay off if their governance holds up to the same scrutiny as the science. Weights are IP. Gradients are regulated data. Errors have causes that live behind firewalls the coordinator cannot see. And regulators do not want to hear “the consortium is responsible”; they want a sponsor of record with a validation file. Every FL agreement should reflect that reality on its first page, not its last.

Sakara Digital works with pharma and biotech organizations building federated learning consortia and the governance structures they run on. If you are exploring an FL collaboration, drafting a consortium agreement, or trying to figure out how to allocate regulatory exposure across participants, and want an independent perspective on where to start, we are happy to have that conversation.