Machine Learning for Red Team Hackers

Learn from scratch to build a Selenium bot that uses machine learning to navigate a website and attempt to capture a captcha, with a hands-on assignment.

Develop machine learning to read captchas on a website, build a dataset by saving captures and training a classifier, and inspect the really simple capture plugin’s source to understand generation.

Construct a WordPress plugin to generate a huge dataset of captures by looping the really simple capture tool, saving images for neural network classification.

Preprocess the captcha dataset by labeling captures, converting to grayscale, thresholding, and cropping character images from contours, then sort by x and save for neural network training.

Train a deep convolutional neural network to classify grayscale character images by normalizing to 20x20, padding to square, and preparing one-hot labels; perform 80/20 train-test split and train with Adam.

Learn how to build a captcha breaking bot using Selenium, a Chrome web driver, and a CNN-based character predictor, including loading models, preprocessing, and automated submission.

Explore fuzzing as an automated software testing technique, using random inputs to trigger crashes, and harness machine learning to make fuzzing more targeted and efficient, with AFL as a reference.

Use a toy fuzzer to generate random strings of configurable length and character sets for a CGI decoder. Track and visualize code coverage to reveal untested lines and guide mutations.

Explore efficient fuzzing through byte mutations; replace, delete, insert, and swap—then chain mutations with code coverage and evolution toward discovering vulnerabilities, in an AFL-style buzzer.

Explore evolutionary fuzzing with genetic algorithms, evolving fuzzing inputs in a population using fitness, mutation, and crossover to maximize code coverage and reveal vulnerabilities.

Fuzz a vulnerable program with AFL by compiling with AFL, seeding inputs, and running mutations to reveal crashes. Manage seeds, outputs, and crashes, including core dumps, to validate the vulnerabilities.

Learn to evade machine learning malware classifiers and iteratively improve them, comparing Malecon neural network and Ember gradient boosting with hand engineered features on P files in adversarial machine learning.

Explore malware detection with the Ember machine learning classifier in a secure VM, covering installation, gradient boosting, and how model scoring thresholds classify samples.

Learn to read and modify PE files with two libraries, leaf and pefile, compare their syntax, extract headers and timestamps, and set a new entry point before saving.

Explore how adversaries evade the amber machine learning malware classifier by injecting benign signals into jigsaw ransomware, and discuss defensive considerations and limitations of gradient-boosting detectors.

Explore neural network models and script-driven predictions in a red-team hacking context. Examine how replacing model types and running the script influence predictions on malicious samples, including Malakal.

Learn how to attack machine learning models using black-box and white-box adversarial attacks, including FGSM, generate examples, and test on real services to harden model defenses.

Perform white-box attacks on neural networks with Foolbox and the fast gradient sign method (fgsm). Compute input gradients via backpropagation to craft adversarial examples that cause misclassification.

Explore Clarif.ai's model gallery—apparel, celebrity, demographics, face detection, food, and moderation models—and learn to access the API with a Python client after creating a developer account.

Learn how to analyze a black-box attack on Clarif.AI moderation models and craft adversarial images through guided perturbations to alter classifier predictions. Explore model choice, scoring, and iterative testing.

Craft adversarial instances from a given image to attack a trained model and trigger misclassification responses, exploring practical adversarial machine learning techniques.

Explore deepfake technology powered by neural networks, transferring faces, voices, and styles across videos and images, and learn to set up, train, and create a converter prediction on device.

Execute a dry run of a deepfake face swap using a gui workflow, from face extraction and alignment to lightweight model training, prediction, and result preview.

Learn to set up a gpu-enabled cloud rig for face swap, installing the Google Cloud SDK and a VNC viewer to run and monitor deepfake training.

Create a deep fake for the assignment, using any source and target, and upload it to YouTube so we can see what you've created.

Examine how attackers steal machine learning models by emulating black-box APIs, extracting probabilities and gradients, and creating surrogate models for adversarial attacks and confidential training-data leakage.

Demonstrate backdoor attacks on machine learning by training data tampering that inserts hidden behavior, illustrated through a facial recognition model using one-shot learning and a covert user with root access.