Implementation of papers with real-time visualizations and parameter control.
From one of the first papers on Adversarial examples - Explaining and Harnessing Adversarial Examples,
The direction of perturbation, rather than the specific point in space, matters most. Space is not full of pockets of adversarial examples that finely tile the reals like the rational numbers.
This project examines this idea by testing the robustness of a DNN to randomly generated perturbations.
$ python3 explore_space.py --img images/horse.png
This code adds to the input image (img
) a randomly generated perturbation (vec1
) which is subjected to a max norm constraint eps
. This adversarial image lies on a hypercube centerd around the original image. To explore a region (a hypersphere) around the adversarial image (img + vec1
), we add to it another perturbation (vec2
) which is constrained by L_{2} norm rad
.
Pressing keys e
and r
generates new vec1
and vec2
respectively.
The classifier is robust to these random perturbations even though they have severely degraded the image. Perturbations are clearly noticeable and have significantly higher max norm.
horse | automobile | : truck : |
In above images, there is no change in class labels and very small drops in probability.
A properly directed perturbation with max norm as low as 3, which is almost imperceptible, can fool the classifier.
horse | predicted - dog | perturbation (eps = 6) |
$ python3 fgsm_mnist.py --img one.jpg --gpu
$ python3 fgsm_imagenet.py --img goldfish.jpg --model resnet18 --gpu
fgsm_mnsit.py
- for attack on custom model trained on MNIST whose weights are 9920.pth.tar
.
fgsm_imagenet
- for pretrained imagenet models - resnet18, resnet50 etc.
epsilon
(max norm)esc
- closes
- save perturbation and adversarial imageAdversarial Image | Perturbation |
---|---|
Pred: 4 | eps: 38 |
Pred: 7 | eps: 60 |
Pred: 8 | eps: 42 |
Pred: 8 | eps: 12 |
Pred: 9 | eps: 17 |
Paper: Adversarial examples in the physical world
$ python3 iterative.py --img images/goldfish.jpg --model resnet18 --target 4
# If argument 'target' is not specified, it is untargeted attack
epsilon
(max norm of perturbation) and iter
(number of iterations)esc
close and space
to pauses
save perturbation and adversarial imageExistence of single pixel adversarial perturbations suggest that the assumption made in Explaining and Harnessing Adversarial Examples that small additive perturbation on the values of many dimensions will accumulate and cause huge change to the output, might not be necessary for explaining why natural images are sensitive to small perturbations.
$ python3 one_pixel.py --img airplane.jpg --d 3 --iters 600 --popsize 10
d
is number of pixels to change (L_{0} norm)
iters
and popsize
are paprameters for Differential Evolution
Attacks are typically successful for images with low confidence. For successful attacks on high confidence images increase d
, i.e., number of pixels to perturb.
bird [0.8075] | deer [0.8933] | frog [0.8000] | bird [0.6866] | deer [0.9406] |
Paper | IJCAI 2018
$ python3 advgan.py --img images/0.jpg --target 4 --model Model_C --bound 0.3
Each of these settings has a separate Generator trained. This code loads appropriate trained model from saved/
directory based on given arguments. As of now there are 22 Generators for different targets, different bounds (0.2 and 0.3) and target models (only Model_C
for now).
$ python3 train_advgan.py --model Model_C --gpu
$ python3 train_advgan.py --model Model_C --target 4 --thres 0.3 --gpu
# thres: Perturbation bound
Use --help
for other arguments available (epochs
, batch_size
, lr
etc.)
$ python3 train_target_models.py --model Model_C
For TensorBoard visualization,
$ python3 generators.py
$ python3 discriminators.py
This code supports only MNIST dataset for now. Same notations as in paper are followed (mostly).
There are few changes that have been made for model to work.
ReLU
on the last layer. If input data is normalized to [-1 1] there wouldn't be any perturbation in the negative region. As expected accuracies were poor (~10% Untargeted). So ReLU
was removed. Also, data normalization had significat effect on performance. With [-1 1] accuracies were around 70%. But with [0 1] normalization accuracies were ~99%.pert
) and adversarial images (x + pert
) were clipped. It's not converging otherwise.These results are for the following settings.
ReLU
at the end in Generatorstep_size
5, gamma
0.1 and initial lr
- 0.001Target | Acc [thres: 0.3] | Acc [thres: 0.2] |
---|---|---|
Untargeted | 0.9921 | 0.8966 |
0 | 0.9643 | 0.4330 |
1 | 0.9822 | 0.4749 |
2 | 0.9961 | 0.8499 |
3 | 0.9939 | 0.8696 |
4 | 0.9833 | 0.6293 |
5 | 0.9918 | 0.7968 |
6 | 0.9584 | 0.4652 |
7 | 0.9899 | 0.6866 |
8 | 0.9943 | 0.8430 |
9 | 0.9922 | 0.7610 |
Pred: 9 | Pred: 3 | Pred: 8 | Pred: 8 | Pred: 4 | Pred: 3 | Pred: 8 | Pred: 3 | Pred: 3 | Pred: 8 |
Target: 0 | Target: 1 | Target: 2 | Target: 3 | Target: 4 | Target: 5 | Target: 6 | Target: 7 | Target: 8 | Target: 9 |
---|---|---|---|---|---|---|---|---|---|
Pred: 0 | Pred: 1 | Pred: 2 | Pred: 3 | Pred: 4 | Pred: 5 | Pred: 6 | Pred: 7 | Pred: 8 | Pred: 9 |
Pred: 0 | Pred: 1 | Pred: 2 | Pred: 3 | Pred: 4 | Pred: 5 | Pred: 6 | Pred: 7 | Pred: 8 | Pred: 9 |
Pred: 0 | Pred: 1 | Pred: 2 | Pred: 3 | Pred: 4 | Pred: 5 | Pred: 6 | Pred: 7 | Pred: 8 | Pred: 9 |
Paper | ICLR 2018
Refer View Synthesis by Appearance Flow for clarity.
$ python3 stadv.py --img images/1.jpg --target 7
Requires OpenCV for real-time visualization.
Column index is target label and ground truth images are along diagonal.