Title: Unseen Risks in the Federated Personalization of Foundation Models

Foundation models are increasingly being tailored to decentralized, private datasets via federated learning, a trend now scaling up amid tightening regulatory demands for post-market surveillance. This convergence gives rise to a specific, yet largely overlooked, category of trustworthiness breakdowns, which we label "Silent Failures." These issues—ranging from exacerbated bias and fairness degradation to the erosion of alignment—are particularly insidious because the privacy safeguards inherent in federated learning obscure visibility into model behavior, making these failures hard to spot.

A review of current benchmarks highlights a structural gap. While federated benchmarks focus on system-level performance, they offer little insight into internal model dynamics. Conversely, centralized trustworthiness benchmarks do evaluate behavior but demand model access that violates federated privacy principles. To address this, we propose a taxonomy detailing six distinct silent failure modes resulting from the interplay between foundation model personalization, dataset shifts, and fundamental federated constraints. Our findings indicate that relying solely on privacy-preserving training methods does not guarantee trustworthy deployment. We conclude by outlining a research agenda focused on behavioral evaluation that respects privacy, advocating for the recognition of silent failures as a standard diagnostic metric for secure and trustworthy federated artificial intelligence.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC