Silent Failures in Federated Personalization of Foundation Models

The paper introduces the concept of "Silent Failures" in the federated personalization of foundation models, highlighting issues like amplified bias and alignment erosion that are difficult to detect due to privacy constraints in federated learning. It presents a taxonomy of six failure modes linked to the interplay between foundation model personalization and dataset shifts, emphasizing that current benchmarks inadequately assess model behavior in a federated context. This work underscores the necessity for new methodologies in privacy-preserving behavioral evaluation to enhance the trustworthiness of federated AI deployments.