Secure Federated Learning: Enhancing Privacy In Collaborative Machine Learning Environments
Year: 2026
Type: conference-paper
Abstract: As massive data is becoming generated at distributed sources, and the importance of user privacy and data security grows, a new type of machine learning, Federated Learning, allows multiple clients to jointly learn a common global model but does not transfer the training data to the server (non-IID data). But FL is not secure by nature considering its intrinsic strength. Even with model updates, gradients, and colluding participants, sensitive information may still be leaked. This raised the necessity of Secure Federated Learning, which is a substantial improvement over FL that combines multiple privacy-preservation and protection methods to ensure the privacy and integrity of data/models, as well as users’ trust in distributed machine learning environments. Secure Federated Learning follows a set of the best mechanisms such as differential privacy, homomorphic encryption, secure multi-party computation and trusted execution environments to ensure the data is kept secure during the training process. In the differential privacy case, the model updates are noised statistically such that the amount of private data that can be recovered is bounded. Using homomorphic encryption allows for operations to be performed on encrypted data, meaning that the data is still secure while being utilized. Secure multi-party computation (SMPC) allows a group of parties to jointly compute a certain function over their inputs without disclosing those inputs. TEEs scrape upon hapless iron and enlist secure radix over which worthy I/Os may be relinquished from the repose of the macrocosm. By way of example, SFL goes beyond data privacy to include communication efficiency, robustness against adversarial attacks, and trust management among heterogeneous clients. Different clients in a federated network may, for instance, differ in their motivation to participate, computational power and trustworthiness. To obscure single updates and ensure privacy information leak, it is adequate to use secure aggregation protocols. To make our model robust against the global model poisoning attempts of malicious clients, we introduce the incorporation of anomaly detection and reputation systems in our model as well. Some use-cases of SFL are applicable to diverse Heterogeneous back end applications like Healthcare, Finance, smart cities, and IoT ecosystems where identity sensitive information(stored here is client specific so) of individuals or organizations needs to remain local.
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7203668816