Fast Preemption: Forward-Backward Cascade Learning for Efficient and Transferable Preemptive Adversarial Defense

Abstract: {Deep learning has made significant strides but remains susceptible to adversarial attacks, undermining its reliability. Most existing research addresses these threats after attacks happen. A growing direction explores preemptive defenses like mitigating adversarial threats proactively, offering improved robustness but at cost of efficiency and transferability. This paper introduces Fast Preemption, a novel preemptive adversarial defense that overcomes efficiency challenges while achieving state-of-the-art robustness and transferability, requiring no prior knowledge of attacks and target models. We propose a forward-backward cascade learning algorithm, which generates protective perturbations by combining forward propagation for rapid convergence with iterative backward propagation to prevent overfitting. Executing in just three iterations, Fast Preemption outperforms existing training-time, test-time, and preemptive defenses. Additionally, we introduce an adaptive reversion attack to assess the reliability of preemptive defenses, demonstrating that our approach remains secure in realistic attack scenarios.