Neural Networks & Deep Learning: Understanding AI for Cybersecurity

Everyone is going crazy about AI. With all the hype, it's easy to forget that it existed long before. Let's delve into neural networks and deep learning to help people in cybersecurity keep up with the latest trends.

"Let's assume we have a thousand layers of networks, but still no idea what to do!"

Popular Architectures: Perceptron Convolutional Neural Networks (CNN) Recurrent Neural Networks (RNN) Autoencoders

Just as human brains have neurons that allow us to transfer data at a 100% rate, AI models transfer about 30% of data at a time. Any neural network is a collection of neurons, essentially a network of inputs and outputs. It's tasked with taking all inputs, performing a function, and sending the result. https://encrypted-tbn0.gstatic.com/images?q=tbn:ANd9GcT9K40eP1YQPH7Ci_pLAFrvXN8fgSL1R37NKA&s

Connections are links between the neurons. They connect multiple outputs of one neuron with inputs from another. Think of them as signals.

Just like we train babies, we need to train the model. We add multiple layers to achieve a desired result.

The network activates associated neurons, which in turn activate the next layers, and so on, until the final result is achieved.

Remember, it's not about achieving a 100% result, but reducing errors with each iteration. When we program these models, everything is matrices, and calculations are based on matrix multiplication. Once a good network is constructed, we assign proper weights for the neurons to react appropriately to the inputs. This involves many cycles of testing and adjusting, a process known as Backpropagation.

Key Differences: The primary difference between deep learning and classical neural networks is in the training capabilities that handle larger networks.

Popular deep learning libraries include: Keras TensorFlow PyTorch

Even with a mini-network of five layers, these libraries offer powerful tools for building robust models.

Some Popular Networks:

Convolutional Neural Networks (CNN)

Convolutional Neural Networks (CNNs) are highly regarded for their ability to identify objects in photos and videos. Extracting features from images is challenging because each image must be manually labeled to teach the machine where specific features, like a cat's ears or tail, are located.

Initially, handcrafted features were used, which meant that if a cat's ears were down, the algorithm wouldn't recognize them. Explaining this to a machine wasn't feasible. Instead, the machine needs to learn these features independently.

We divide the entire image into pixel blocks and assign each block a dominant line type. This process is known as convolution, where each layer of the neural network acts as a different function. By inputting our neural network with numerous cat photos, it assigns combinations of lines it has seen. The network then searches for the most distinctive features of the objects autonomously.

Recurrent Neural Networks (RNN)

These models are used in speech recognition and voice synthesis in smart devices (handling speech, text, and music). They are trained to speak not letter by letter, but in whole lines. By using multiple voice samples, we train a neural network to generate audio, continuously minimizing errors with each iteration.

These models consist of consecutive units like syllables, which depend on the previous ones. For instance, in models like Jarvis, they remember previous responses by adding memory to each neuron and use it as additional input in the next run. Limiting the neurons is crucial because the data can become large; thus, it's essential to restrict it to the most recent n results.

Are we in a terminator movie ?

We won't get a Skynet! There is a certain scale of intelligence based on the data provided. Even elephants remember and process more data and are intelligent. AI will help us remember and process even more information. Don't worry security professionals just catch attackers with it.

Want to learn more I recommend Reference