Explore our new Cybersecurity Terms Glossary, explained by our experts.
Cybersecurity Terms Glossary
Federated learning trains a shared model without pooling the data, but the updates participants exchange can leak their training records and can be crafted to poison the global model.

What Is Federated Learning Security?
Federated learning trains a shared model across many parties without any of them sending their raw data anywhere. Each participant trains locally on its own records and sends only a model update; a server aggregates those updates into a new global model and redistributes it. Google deployed the approach for mobile keyboard prediction in 2017, and it has since become the standard proposal whenever data cannot legally or commercially be pooled—between hospitals, between banks sharing fraud signals, or across devices.
Cross-device federation involves thousands or millions of phones and sensors, most of them untrusted, each contributing a tiny amount. Cross-silo federation involves a handful of institutions with their own infrastructure, contracts, and legal teams. The security question differs accordingly: at device scale the realistic threat is a population of malicious clients, while in a silo arrangement the questions are whether a partner can reconstruct your data from your updates and whether the aggregating party is trustworthy.
The central assumption—that gradients reveal nothing about the data—does not hold. Work published in 2019 under the title Deep Leakage from Gradients showed that original training images and labels could be reconstructed pixel by pixel from the gradients of a small batch. Later research extended reconstruction to larger batches and to text, and showed that a malicious server can modify the model it distributes to make reconstruction easier. Related attacks infer whether a specific record was in a participant’s dataset, which is the membership inference problem arriving through a side door, or infer properties of a participant’s population, such as the demographic mix of a hospital’s patients.
Because participants train locally, nobody inspects their data or their process. A malicious client can send crafted updates to degrade the global model or to implant a backdoor that fires on a chosen trigger while leaving normal accuracy untouched—a scenario demonstrated in How to Backdoor Federated Learning in 2020. Sybil attacks make it worse: in a cross-device setting, one attacker controlling many fake clients can outweigh honest contributions. This is ordinary data poisoning, but with the defender denied the usual remedy of inspecting the training set.
Federated learning is frequently presented as a way to avoid data protection obligations, and that claim needs care. Model updates derived from personal data may themselves be personal data where reconstruction is feasible, so the ICO’s identifiability test applies to the updates, not just to the raw records. The approach is nonetheless strong evidence of data protection by design under Article 25 and of data minimisation, and the ICO treats federated analytics as a privacy-enhancing technology in its guidance. In practice the arrangement also needs clear allocation of controller and processor roles across participants, a joint controller agreement where decisions are shared, and a documented security assessment of the aggregation protocol. In health research the comparable UK pattern is the secure data environment, where analysts come to the data rather than the data moving; federated learning fits the same policy direction. At Privia Security, our Cyber Risk Analysis Services assess distributed training architectures before partners commit to them.
The raw records do not, which is the main benefit. Model updates do leave, and those updates can leak information about the data that produced them, sometimes enough to reconstruct individual examples. Treat the updates as a controlled export, protect them with secure aggregation and differential privacy, and document that analysis rather than relying on the raw-data argument alone.
Combine prevention with detection. Authenticate and, where possible, attest clients so fake participants are expensive to create; clip and bound each update so no single contributor can dominate a round; use robust aggregation to discard outliers; and monitor contributions across rounds for participants who consistently pull the model in an unusual direction. None of these is sufficient alone against a determined, patient attacker.
No. It prevents the server from seeing any individual update, which closes the most direct reconstruction path, but the aggregate itself still carries information, and a colluding majority of clients can isolate a victim’s contribution. Pair it with differential privacy for a formal bound, and with robust aggregation to address poisoning, which secure aggregation does nothing to prevent.
When the data genuinely cannot be centralised for legal, contractual, or volume reasons, and the participants share an interest in one model. If you could lawfully pool the data in a single controlled environment, that is usually simpler, cheaper, and easier to secure. Federation adds a distributed protocol, several new attack surfaces, and significant engineering overhead.
You May Be Interested In These
We use a few cookies that are needed to make this website work, and they are always on. We would also like to set analytics cookies, to see which pages are useful; functional cookies, so that our live chat works; and advertising cookies, to measure our campaigns. We will not set any of those until you say yes, and you can change your mind at any time from the Cookie Settings link in the footer.
Choose the cookies you are happy for us to use. Strictly necessary cookies are always on; everything else is off until you turn it on.
These cookies keep the website working: they hold your security checks, your form submissions and the choice you make here. The site cannot run without them, so they are always on.
These let our live chat open and remember your conversation, and let the newsletter sign-up work. Switch them off and the chat window will not appear.
These tell us which pages are read and how people find us, so that we can improve the site. The reports are about groups of visitors, never about you personally.
These let us measure our advertising and show you our adverts on other websites, such as LinkedIn, Google and Meta. Switch them off and you will still see adverts, but they will be less relevant.