arXiv:2610.19077
Federated Unlearning of Quantum Data via Classical Shadows
Zihao Ding, Jun Huang, Cong-cong Xing
Abstract
Quantum federated learning trains a quantum neural network (QNN) across nodes that hold quantum data, such as the states of quantum sensors. To forget data, unlearning methods for quantum models replay them, and federated unlearning from stored updates removes whole nodes only. Quantum data cannot be replayed, since a state is consumed when measured and cannot be cloned. We propose Shadow-based Data Erasure (SHADE), in which each node records a classical shadow of every training state, a few randomized single-qubit measurements taken at training time. The gradient and curvature of a QNN on a state are expectations of observables, so SHADE estimates them for the forget data from the shadows at any later parameters and removes their effect by a federated Newton correction. Only classical estimates cross the network. We prove that the shadow size needed for a target unlearning error grows exponentially in the size of the backward light cone of the measured observable and not in the number of qubits. In simulation with 50 nodes and on edge boards, SHADE matches the forget accuracy of retraining within 0.5 points, without replay and in under 0.2 s.
AI Summary
The paper addresses deletion of selected training states from a federated quantum neural network when neither forgotten nor retained quantum data can be prepared again. SHADE allocates training-time copies of each state to local-Pauli classical shadows, storing only qubits in the output observable’s backward light cone. At deletion time, nodes estimate the forgotten data’s gradient and retained data’s Hessian at the current parameters; the server applies one Newton correction, and nodes erase the forgotten states’ shadows. Only classical estimates are communicated. For parameters independent of the snapshots, shadow averages give unbiased estimates of circuit outputs and their first and second derivatives. A median-of-means argument bounds the snapshots per state by $S=O(4^{|\Lambda|}\varepsilon^{-2}\log(|\mathcal D|(P+1)^2/\delta))$ for simultaneous additive error $\varepsilon$. Under local strong convexity and a Lipschitz Hessian, a parameter-distance bound separates Newton-step approximation error from gradient and Hessian estimation errors. Experiments simulate 50 nodes using eight-qubit Ising/XXZ ground-state data and a 24-parameter circuit. With 32,000 snapshots per state, SHADE reports forget-set accuracy within 0.5 percentage points of retraining and correction times below 0.2 seconds, storing 72 kB per state. Edge-board tests also simulate quantum operations, using a device-derived noise model. The stated limitations include exponential storage and classical computation in light-cone size, finite-copy constraints, and a Newton approximation error that more snapshots cannot eliminate. Sequential requests require fresh snapshot groups to preserve independence; the retraining-distance guarantee assumes strong convexity in the relevant region.
Error found: The abstract's target-unlearning-error claim is conditional on optimization and curvature assumptions and on the target exceeding an irreducible one-step Newton error. Table I reports nonzero distances for Retrain despite defining those metrics as distances to Retrain, without specifying a separate reference retraining run. Table I assigns Retrain hundreds of millions of forgotten-state replay copies even though the described retraining procedure uses only retained states.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.