adversarial-training

name	adversarial-training
version	2.0.0
description	Defensive techniques using adversarial examples to improve model robustness and security
sasmp_version	1.3.0
bonded_agent	05-defense-strategy-developer
bond_type	PRIMARY_BOND
input_schema	{"type":"object","required":["training_method"],"properties":{"training_method":{"type":"string","enum":["standard","trades","certified","ensemble","all"]},"epsilon":{"type":"number","default":0.3},"attack_types":{"type":"array","items":{"type":"string","enum":["fgsm","pgd","cw","autoattack"]}}}}
output_schema	{"type":"object","properties":{"robustness_score":{"type":"number"},"clean_accuracy":{"type":"number"},"adversarial_accuracy":{"type":"number"}}}
owasp_llm_2025	["LLM04","LLM09"]
nist_ai_rmf	["Manage"]

Adversarial Training

Build robust AI models by training with adversarial examples and attack simulations.

Quick Reference

Skill:       adversarial-training
Agent:       05-defense-strategy-developer
OWASP:       LLM04 (Data Poisoning), LLM09 (Misinformation)
NIST:        Manage function
Use Case:    Improve model robustness against attacks

Training Methods

1. Standard Adversarial Training

Method: standard
Robustness Gain: 30-50%
Accuracy Tradeoff: 5-15%
Complexity: Medium

class AdversarialTrainer:
    def __init__(self, model, epsilon=0.3, attack_steps=10):
        self.model = model
        self.epsilon = epsilon
        self.attack_steps = attack_steps

    def train_step(self, x, y):
        # Generate adversarial examples using PGD
        x_adv = .pgd_attack(x, y)

        
        loss_clean = .criterion(.model(x), y)
        loss_adv = .criterion(.model(x_adv), y)

        
        total_loss =  * loss_clean +  * loss_adv
         total_loss

     ():
        
        x_adv = x.clone().requires_grad_()

         _  (.attack_steps):
            loss = .criterion(.model(x_adv), y)
            loss.backward()

            
            x_adv = x_adv + .epsilon/.attack_steps * x_adv.grad.sign()
            
            x_adv = torch.clamp(x_adv, x-.epsilon, x+.epsilon)
            x_adv = x_adv.detach().requires_grad_()

         x_adv