Shipped: Black-Box Neural Network Hardening — AI Defense Framework

An open-source adversarial AI security framework that investigates how to strengthen black-box neural networks without access to their internal weights, architecture, or original training data.

The seven-stage experimental pipeline uses WGAN-GP generative models to synthesize training data, knowledge distillation to create high-fidelity surrogate neural networks, and adversarial training to develop defensive capabilities. The framework evaluates attacks including FGSM, PGD, Carlini & Wagner, AutoAttack, BPDA, and EOT, alongside layered defenses such as JPEG compression, bit-depth reduction, and denoising.

Evaluated across MNIST, CIFAR-10, and TinyImageNet, the project reports surrogate agreement rates of 94.7%, 85.4%, and 78.0%, respectively. The research explores how surrogate-based defense strategies can improve adversarial robustness in environments where direct model modification is restricted.

Designed for AI security researchers, machine learning engineers, and organizations investigating the protection of proprietary AI systems.