Privacy-Preserving Active Learning for heritage language revitalization programs across multilingual stakeholder groups

Privacy-preserving active learning is a groundbreaking approach that could revolutionize heritage language revitalization programs. In a recent project, the author encountered a critical problem that no academic paper had prepared them for - the question of where people's voices go when uploaded to a digital platform. This raised complex issues of data sovereignty, consent, and trust among multilingual stakeholders.

The true constraints in these programs are governance issues like data ownership and consent. Privacy-preserving active learning emerged as a solution to these challenges, focusing on privacy as a fundamental aspect of the system rather than an afterthought. Heritage languages face extreme data scarcity, limited annotation resources, and distrust from various stakeholders. Add multilingual complexities to the mix, and standard active learning techniques fall short.

The author devised a layered architecture to handle these unique demands. Federated learning keeps raw data on community devices; only model updates are transmitted. Differential privacy bounds the information revealed in those updates, preventing any individual's contribution from being traced back to them. Secure aggregation ensures the server cannot inspect individual updates at all.

A key aspect of the approach is rethinking the query strategy in federated active learning. Standard active learning selects the most uncertain samples from a central pool, but when the pool is distributed among clients, the server must access client uncertainty scores instead. These scores are calculated locally and then aggregated without sharing raw data.

To maintain privacy, the author applied differential privacy techniques per client. Client-level DP was used, where each client's update is clipped and then injected with calibrated Gaussian noise. This approach balances privacy protection and model learning efficiency. Pairwise masking secure aggregation was employed to ensure the server never sees individual updates.

In summary, privacy-preserving active learning offers a promising solution to the complex problem of heritage language revitalization. By addressing data sovereignty, consent, and trust concerns while incorporating principles of federated learning, differential privacy, and secure computation, this approach could enable communities to retain control over their language data while still benefiting from AI technologies.