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

Deep learning has advanced significantly but remains vulnerable to adversarial attacks, compromising its reliability. While conventional defenses typically mitigate perturbations post-attack, few studies explore proactive strategies for preemptive protection. This paper proposes Fast Preemption, a novel defense that neutralizes adversarial effects before third-party attacks. By employing distinct models for input labeling and feature extraction, Fast Preemption enables an efficient, transferable preemptive defense with state-of-the-art robustness across diverse systems. To further enhance efficiency, we introduce a forward-backward cascade learning algorithm that generates protective perturbations, leveraging forward propagation for rapid convergence and iterative backward propagation to mitigate overfitting. Executing in just three iterations, Fast Preemption surpasses existing training-time, test-time, and preemptive defenses. Additionally, we propose the first effective white-box adaptive reversion attack to evaluate the reversibility of preemptive defenses, demonstrating that our approach remains secure unless the backbone model, algorithm, and settings are entirely compromised. This work establishes a new paradigm for proactive adversarial defense.