VAST-FCL: Verifiable Sparse-Task Federated Continual Unlearning for Cross-Institutional Educational AI

Abstract

Cross-institutional educational models must continually absorb new curricula, semesters, and device streams while honoring requests to withdraw previously authorized data. Continual learning seeks to prevent forgetting, whereas machine unlearning requires the deliberate removal of prior influence. This creates a structural conflict that is not resolved by deleting database records alone. The supplied course architecture uses task-specific sparse masks and parameter resets to isolate knowledge, but its claim of complete forgetting conflates three different objectives: parameter removal, distributional similarity to retraining, and externally verifiable execution. This paper proposes VAST-FCL, a Verifiable Audited Sparse-Task Federated Continual Learning framework. VAST-FCL confines revocable task increments to addressable parameter subspaces, repairs downstream tasks through controlled mask reallocation and authorized rehearsal, and records each state transition in a hash-linked transparency log. We evaluate the method on six sequential synthetic educational tasks with 72 features and 12 random seeds. After deleting the third task, VAST-FCL retains 79.07% ± 1.79% mean accuracy, compared with 77.19% ± 2.20% for reset-only and 79.73% ± 1.57% for full retraining. The gain over reset-only is significant (t(11) = 5.31, p < 0.001), while estimated computation is 22% of full retraining. All protocol assertions pass across the 12 runs. The findings support efficient task-level structural deletion, but not universal sample-level distributional equivalence. Production deployment therefore requires a layered evidence package that combines structural checks, behavioral privacy tests, and independent audit.