Quantum Annealing Boosts Byzantine-Resilient Federated Learning
A recent study published on arXiv introduces a quantum annealing technique aimed at enhancing Byzantine resilience in federated learning. This approach redefines client selection as a Quadratic Unconstrained Binary Optimization (QUBO) challenge, which quantum annealers can solve. In contrast to the greedy MultiKrum algorithm, QUBO optimally evaluates all client subsets to identify the closest group. Tests involving 15 clients demonstrated that QUBO surpassed MultiKrum in the face of difficult attacks: the accuracy for Advanced LIE detection was 95.11% compared to MultiKrum's 81.33% on MNIST, and 97.78% versus 75.56% on CIFAR-10. This method tackles weaknesses in federated learning where harmful updates can mimic the statistical traits of legitimate ones.
Key facts
- Federated Learning trains a global model across decentralized clients while preserving data privacy.
- Byzantine-resilient aggregation methods like MultiKrum score gradients against nearest neighbors.
- MultiKrum can miss malicious updates that preserve statistical properties of honest ones.
- The proposed method uses quantum annealing to reformulate client selection as a QUBO problem.
- QUBO encodes pairwise distances into a cost function solved by quantum annealers.
- QUBO jointly optimizes over all subsets to find the mutually closest group of m clients.
- At 15 clients, QUBO outperformed MultiKrum on Advanced LIE attacks.
- Advanced LIE detection accuracy: 95.11% vs 81.33% on MNIST, 97.78% vs 75.56% on CIFAR-10.
Entities
Institutions
- arXiv