Biased Boundary Attack: ImageNet implementation

This repository contains an implementation of the Boundary Attack on ImageNet, which is accelerated by various priors (Perlin/low frequency, regional masking, surrogate gradients).

• Paper on arXiv: Guessing Smart: Biased Sampling for Efficient Black-box Adversarial Attacks
• Our winning submission to the NeurIPS 2018 Adversarial Vision Challenge was based on this work: https://github.com/ttbrunner/biased_boundary_attack_avc
• Click the image below for an animated demonstration!

https://www.youtube.com/watch?v=ljUmA66Gr0A&list=PL4qRc8jGTvrCpjKSmnKXRCHmsIrMKktX2

Usage:

• Download model checkpoints to models/, or use your own models
• Enable or disable biases in bench_settings.py
• Set path to ImageNet in run_imagenet_bench.py
• Start run_imagenet_bench.py
• Images are logged to "./out_imagenet_bench/".