Adversarial Robustness Toolbox for AI Model Poisoning | MNIST Backdoor Attack Tutorial

In this AI security experiment, we use the Adversarial Robustness Toolbox (ART), PyTorch, and the MNIST handwritten-digit dataset to demonstrate a controlled model poisoning and backdoor attack. The surprising part is that the poisoned model still looked healthy during normal testing. 📊 Our recorded results: Clean model accuracy: 96.5% Poisoned model clean accuracy: 95% Training samples poisoned: 1,800 of 12,000 — 15% Clean model trigger target rate: 0.39% Poisoned model attack success rate: 100% in this controlled run That means normal accuracy dropped by only 1.5 percentage points, yet the poisoned model behaved completely differently when the backdoor trigger appeared. The experiment demonstrates an important AI security lesson: High model accuracy does not automatically mean the model is secure. 🔐 A compromised model can continue behaving normally under ordinary conditions while containing hidden behavior that only activates under specific inputs. In this tutorial, we walk through the experiment from scratch, including: 🧠 Building the MNIST neural network with PyTorch 📦 Loading and preparing the MNIST dataset 🧪 Creating a controlled backdoor using ART ☣️ Poisoning part of the training data 🏋️ Training clean and poisoned models 🎯 Testing the hidden trigger 📊 Comparing clean accuracy and attack success 💾 Saving the models, metrics, charts, and experiment evidence This experiment is performed in a controlled lab environment for authorized AI security research, education, and defensive testing. The 100% attack success rate shown in this video is the result of this specific experiment and should not be interpreted as a universal result for every model, dataset, poison rate, or backdoor configuration. 🛠️ Tools & Resources Used Adversarial Robustness Toolbox (ART) https://github.com/Trusted-AI/adversa... ART Documentation https://adversarial-robustness-toolbo... PyTorch https://pytorch.org/get-started/locally/ Torchvision MNIST Dataset https://docs.pytorch.org/vision/main/... Torchvision GitHub https://github.com/pytorch/vision JupyterLab https://jupyter.org/install Python https://www.python.org/ NumPy https://numpy.org/ pandas https://pandas.pydata.org/ Matplotlib https://matplotlib.org/ uv Python Package Manager https://docs.astral.sh/uv/ Parrot OS https://parrotsec.org/ Oracle VirtualBox https://www.virtualbox.org/ 🌐 Follow Kiso Secure ▶️ YouTube    / @kisosecure   𝕏 https://x.com/KisoSecure 📸 Instagram   / kisosecure   💼 LinkedIn   / kiso-secure   At Kiso Secure, we explore offensive AI security, adversarial machine learning, AI red teaming, prompt injection, AI agent security, model and data-pipeline attacks, and practical approaches to securing AI systems. Subscribe for more hands-on AI security experiments, adversarial ML labs, AI red teaming tutorials, and defensive testing. 🔐🤖 #AISecurity #AdversarialMachineLearning #AIRedTeaming