Course Description

Advanced topics in building and verifying software systems, selected from areas of current research such as: model checking and automated verification, testing and automated test generation, program synthesis, runtime verification, machine learning and its applications in the design of verified systems, formal analysis of machine learning algorithms, principles of programming languages and type systems.

Source

HW1 Adversarial Attacks & Training

Adversarial attacks and training on neural networks. Implementation of the Projected Gradient Descent attack described in the Towards Deep Learning Models Resistant to Adversarial Attacks paper.

HW2 Interval Analysis for Neural Networks

HW4

adversarial_explanations/: SmoothGrad explanations for adversarial examples.

explanations/: LIME and SmoothGrad explanations for normal examples.

output: correlation scores, adversarial images

HW5

Problem 1

Problem 1 Google Colab Notebook

To run

Activate your hugging face token

export HUGGINGFACE_TOKEN=your_token

Problem 2

Not fully working. Data loading and model training functionality is present.

Problem 2 Google Colab Notebook

Problem 3

To run

Set your OpenAI API key in a file key.txt in the same directory as the notebook.

Download the VQA Balanced Real Images questions, images, and complimentary pairs from the VQA Balanced Real Images page.

Place the downloaded files in the data directory.

Unzip the files in the data directory.

Problem 3 Google Colab Notebook