Privacy-Preserving Continual Learning for Detecting Concept Drift in Distributed Artificial Intelligence Systems
Abstract
Abstract Distributed artificial intelligence systems increasingly process data streams whose statistical properties change over time, a phenomenon known as concept drift. In centralised pipelines, researchers typically diagnose drift by inspecting raw observations. In federated and edge deployments, however, regulatory and contractual constraints, along with significant re-identification risks associated with sharing gradients and statistics, preclude such inspection. This article introduces and evaluates a framework for privacy-preserving continual learning that detects and adapts to concept drift without exposing client-level data or unsecured model updates. The framework comprises three components: a differentially private drift-signalling channel that transmits perturbed loss and confidence summaries from clients; a server-side adaptive windowing detector that operates on securely aggregated signals; and a regularised rehearsal mechanism that mitigates catastrophic forgetting during adaptation. We evaluate the proposed design on three streaming benchmarks: energy demand, synthetic rotating hyperplanes, and network intrusion traffic, each distributed across fifty non-identically distributed clients under abrupt, gradual, and recurring drift scenarios. Results indicate that drift is detectable at moderate privacy budgets, with a mean detection delay increase of approximately thirty per cent compared to a non-private baseline. Furthermore, accuracy retention on previously…
Read the paper
Related studies
- Active Learning during Federated Learning for Object Detection
- Fairness-aware federated learning
- Decentralized Federated Learning
- Building a Personally Identifiable Information Recognizer in a Privacy Preserved Manner Using Automated Annotation and Federated Learning
- Preservation of Patient Level Privacy: Federated Classification and Calibration Models
- Efficient explainable federated learning
Go deeper in Evidence Hub
Evidence Hub adds a study data sheet (design, sample size, evidence level, key numbers), an AI summary and a chat grounded in the paper's text.