AI and ML are often used interchangeably. Let’s try to provide some simple definitions and examples to understand their relationship and how they fit into our work of defending AI from adversarial attacks.

AI is a field in computer science that involves techniques and approaches to creating intelligent machines and applications that can perform tasks with intelligence normally associated with humans. These tasks include understanding natural language and images, recognizing patterns, solving problems, and making decisions.

AI is integrated with applications and systems. In everyday life, we use AI for things such as predictive texting, email spam filters, and recommendations. With its constant progress, AI can be found in smart homes in Internet of Things (IoT) devices such as security cameras, doorbells, vacuum cleaners, and digital assistants such as Siri or Alexa. Autonomous cars and smart medical devices are other examples of using AI to create...

Depending on how models learn, ML can be classified into three types:

• Supervised learning, where each data sample must have a label indicating the correct outcome. The model learns from labeled structured data, such as CSV files, by adjusting its internal parameters based on its error when it guesses the result. Supervised learning is by far the most used type of learning in classification images, voice and language recognition, numerical forecasting, and more.
• Unsupervised learning, on the other hand, involves training on data, usually unstructured, without labels. Unsupervised learning uses clustering and other techniques to understand the underlying structure of data, identify patterns, and perform anomaly detection, fraud detection, social network analysis, market segmentation, and supervised learning.
• Reinforcement learning relies on an agent to behave in an environment and learn by performing certain actions, observing the results...

Several algorithms in ML have pros and cons and suit different use cases.

In supervised learning, we have the following:

• Linear regression, which predicts a continuous output variable based on input features. It’s used in economics for forecasting and in healthcare for predicting disease progression.
• Logistic regression, which, despite its name, is an algorithm for binary classification problems and estimates the probability an instance belongs to a class. It’s used in credit scoring and medical testing.
• Decision tree, which learns simple decision rules inferred from data features. It’s useful in business decision-making and customer segmentation.
• Random forest, which uses multiple decision trees to prevent overfitting. This makes it an ensemble algorithm and is used in predicting disease risk, loan defaulters, and customer preferences.
• Support vector machine (SVM), which can model complex decision boundaries and...

Inspired by human brain biology, artificial neural networks (ANNs) are good at processing unstructured data such as images, audio, and text and are widely used in image recognition, speech recognition, and natural language processing (NLP). These are their fundamental blocks:

• Neurons and layers: ANNs apply parallel processing by using nodes called neurons. Each node has a weight and a bias, both of which are used to produce their output based on outputs. Neurons are organized in layers, and typically, there is an initial input and final output layer, and layers in between called hidden layers where the actual computation takes place. Inputs to each layer are derived from the outputs of the previous layer.
• Training and weights update: Training an ANN involves adjusting the weights and biases of neurons based on error. This consists of a process called backpropagation and an optimization method, such as batch gradient descent and/or stochastic...

We can develop ML models in many languages, ranging from Python, R, C++, Java, and Julia to scientific tools such as proprietary MATLAB and open source Octave. Python is by far the most widely used language in academia, the scientific community, and industry. This makes it a near de facto standard for mainstream ML development. We will be using Python throughout this book.

Python’s simplicity and readability contribute to its popularity, but what sets it apart is the rich ecosystem of scientific and data analysis libraries. NumPy for numeric computing, pandas for data analysis, and Matplotlib for charting are three libraries that are widely used in ML work.

These are available as standard Python packages. The default packager in Python is pip, and you can find and install packages from package repositories, such as the Python Package Index (PyPI). In some operating systems and environments, you may find pip as pip3.

Packages sometimes have other...

In this chapter, we set the foundations of AI for the rest of this book. We covered some important topics:

• What AI is and its shift toward AGI.
• How ML creates models adaptively by ingesting data and how it is the brain of AI. This makes it the focus of adversarial AI attacks and defenses.
• The different types of ML based on how models learn – that is, supervised, unsupervised, and reinforcement learning.
• The seven typical steps in the ML life cycle, which include data collection and pre-processing, selecting an algorithm based on the problem we are solving, model training, testing and evaluation, fine-tuning and optimization, and, finally, deploying and using the model.
• Key ML algorithms and where they are used. This included linear and logistic regression, decision trees, and their ensemble version with random forests in supervised learning. We looked at K-means clustering and PCA, two popular unsupervised models, and Q-learning in reinforcement...

To learn more about the topics that were covered in this chapter, take a look at the following resources:

• Hands-On Data Preprocessing in Python, by Roy Jafari
• Mastering Machine Learning Algorithms - Second Edition, by Giuseppe Bonaccorso
• Deep Learning with TensorFlow 2 and Keras - Second Edition, by Antonio Gulli, Amita Kapoor, and Sujit Pal