Adversarial Machine Learning: Attacking and Defending Neural Networks

As ML models are deployed in security-critical applications, understanding their vulnerabilities to adversarial manipulation is essential.

Attack Taxonomy

• **Evasion Attacks**: FGSM, PGD, C&W, and AutoAttack — crafting imperceptible perturbations to fool classifiers
• **Poisoning Attacks**: Backdoor injection via data poisoning and model poisoning during federated learning
• **Model Extraction**: Stealing model architecture and parameters through black-box query access
• **Inference Attacks**: Membership inference and attribute inference for privacy violations

Defense Strategies

• **Adversarial Training**: Augmenting training data with adversarial examples (PGD-AT, TRADES)
• **Certified Robustness**: Randomized smoothing for provable L2 robustness guarantees
• **Detection**: Feature squeezing, MagNet, and spectral signature analysis
• **Differential Privacy**: DP-SGD to protect against membership inference