The ML OWASP Top 10
OWASP ML Security Top 10 maps the attack surface of machine learning systems. Here is what each risk means for red teamers and defenders.
Training and evaluating a malware classifier
Training a byteplot CNN on Malimg to 88.54% accuracy, then see why overall accuracy on an imbalanced dataset misleads and what evaluation metrics to use instead
Building a malware classifier on ResNet50
Transfer learning turns a frozen ImageNet backbone into a ResNet50 malware classification model on the Malimg dataset, and shows where that shortcut leaks.
Malware image preprocessing and the accuracy illusion
Malware image preprocessing decides CNN classifier accuracy before training begins. How the Malimg split, resize and normalisation hide the real risk.
Malware as images
Explores the Malimg (Malware) dataset, binary-to-image conversion, and why class imbalance is an adversarial attack surface.
Malware classification
Train a CNN to classify malware families from grayscale binary images using the Malimg dataset, and learn why byte-level texture is both signal and weakness.
Training and evaluation
The latest entry in the AI red teaming series trains a random forest on NSL-KDD and shows how evaluation metrics map the exact weaknesses an attacker exploits.
Preprocessing and splitting the dataset
Preparing the NSL-KDD dataset for random forest anomaly detection, from binary and multi-class targets to encoding, feature selection, and honest splitting.
Network anomaly detection
Train a random forest on the NSL-KDD dataset for network anomaly detection, with every data loading step examined through an adversarial red teaming lens.
Training and evaluating your first spam classifier
Build, tune, and evaluate a Naive Bayes spam classifier with scikit-learn, then examine what the model reveals to an adversary in this AI red teaming entry.