Evade AI Model

Hardened models are more difficult to evade.

Using multiple different models increases robustness to attack.

Using a variety of sensors can make it more difficult for an attacker to compromise and produce malicious results.

Preprocessing model inputs can prevent malicious data from going through the machine learning pipeline.

Prevent an attacker from introducing adversarial data into the system.

Deepfake detection can be used to identify and block generated content.

Execute representative digital, multimodal, and physical-domain evasion attacks. Use successful attacks to improve model robustness, preprocessing, adversarial-input detection, human oversight, and monitoring.