A Reactive Security Framework for Protecting AI Models from Adversarial Attacks: An Autoencoder-Based Approach

This paper proposes a reactive security framework for enhancing the resilience of AI models against adversarial attacks [5, 6, 7, 8]. The framework leverages runtime monitoring, anomaly detection, and model retraining to dynamically adapt to evolving attack strategies. Anomaly detection is performed using an autoencoder-based algorithm that identifies deviations from expected model behavior [8, 9, 10]. Model retraining employs adversarial training to ”immunize” the model against similar attacks [5, 6]. We discuss the choice of autoencoder architectures for different data types and detail the mathematical foundations of both anomaly detection and adversarial training [3]. The framework’s effectiveness is evaluated through simulations and benchmark datasets, demonstrating its ability to secure AI models against diverse adversarial attacks.