Defending against evasion attacks
In the arms race between attackers crafting increasingly sophisticated evasion attacks and defenders bolstering the security of ML models, a multifaceted approach to defense is essential. This section will explore a suite of strategies designed to mitigate the risk and impact of evasion attacks.
Mitigation strategies overview
Defense strategies against evasion attacks can be broadly categorized into reactive and proactive measures. Reactive defenses respond to attacks as they happen, often employing real-time detection and mitigation techniques. On the other hand, proactive measures focus on hardening models against potential attacks even before they occur, such as during the training process or model design phase.
These strategies can also be layered, providing a defense-in-depth approach that secures models at multiple levels. For instance, input preprocessing can be combined with adversarial training to prevent and withstand adversarial...
