Model Hardening

This technique strives to make AI models more robust to adversarial inputs via adversarial training or network distillation. Examples include (i) using randomization to inject noise during training to enhance resilience to evasion attacks (especially triggered by subtle perturbations), (ii) Gradient Masking, (iii) Feature Squeezing.

​

Explanation

Use adversarial robustness techniques to re-train or re-deploy your model securely. 

​

How it works

Reduces the success of the attacker by making the model more resilient to attacks. 

​

How to implement

Adversarial training, network distillation.