On the Fragility of Contribution Score Computation in Federated Learning