Speakers
Description
Black-box adversarial attacks, where the attacker has no access to model gradients, require expensive query-based search over high-dimensional input spaces. To the best of our knowledge, this work for the first time studies the quantum algorithms for this problem and shows the quantum advantage over classical methods. We start from Grover’s algorithm, which offers quadratic improvements for unstructured search, while its practical performance is limited by significant solution-count overhead. To address this, we propose a quantum adversarial search framework based on Grover Adaptive Search (GAS), and provide a theoretical analysis on classical methods, Grover, and GAS to establish conditions under which quantum advantage emerges, along with closed-form computational complexity trade-offs. Empirically, we evaluate our approach on standard image classification datasets spanning from digits to ImageNet against state-of-the-art classical black-box attack and quantum baselines, demonstrating its improved efficiency and increased target coverage. Furthermore, we show that quantum-generated samples can effectively enhance model robustness via closed-loop adversarial training and demonstrate the framework's resilience under NISQ noise. Our results offer both theoretical and empirical evidence that quantum algorithms can accelerate black-box adversarial attacks, opening a new direction at the intersection of quantum computing and machine learning security.
| Topic | Quantum algorithms for machine learning applications |
|---|---|
| Additional Topic | Quantum-enhanced robustness in machine learning models |