Considering Generative Adversarial Networks? A Cyber Intelligence Perspective

Editor’s Note: From time to time, ComplexDiscovery highlights publicly available or privately purchasable announcements, content updates, and research from cyber, data, and legal discovery providers, research organizations, and ComplexDiscovery community members. While ComplexDiscovery regularly highlights this information, it does not assume any responsibility for content assertions.

To submit recommendations for consideration and inclusion in ComplexDiscovery’s cyber, data, and legal discovery-centric service, product, or research announcements, contact us today.

Educational Paper* by Fabio Biondi, Giuseppe Buonocore, and Richard Matthews

Backgrounder: Produced by the NATO Cooperative Cyber Defence Centre of Excellence (CCDCOE) Cyber Intelligence Division, this new paper highlights generative adversarial networks (GAN) as they are an increasingly hot topic in cyber intelligence. GANs are beginning to demonstrate abilities that will assist public intelligence analysts in playing more active roles in global security.

Generative Adversarial Networks from a Cyber Intelligence Perspective

Paper Extract

Generative adversarial networks (GANs) are a deep-learning model first described by Ian Goodfellow in 2014. They use two neural networks – one that creates content and one that analyses it – in a pseudo-game-like adversarial process. To understand what a GAN is, first we must understand some fundamental principles of machine learning. Machine learning is ‘a subset of artificial intelligence (AI) where an artificial intelligence platform is trained to make predictions based on data’. The goal is to build algorithms that can ‘learn’ to make predictions based on new data to which it has access. This data forms training sets from which algorithms can correlate future examples based on prior inputs.

To showcase the power of artificial intelligence, the entire prior paragraph was not written by the authors which is why the definition of machine learning contains no citation. Instead, a neural network was used to complete the passage based on a few seed words (the training set) provided to the network and then manually edited for proof and clarity. Indeed, the first paragraph of the previous section was written in the same manner. This is one power of GAN which has yet to be fully realized; the widespread automation of report writing using GAN tools as the first pass.

In this work, we use the definition of machine learning from Shalev-Shwartz and Ben-David as follows: ‘machine learning refers to the automated detection of meaningful patterns in data’. The common models used in machine learning can be divided into two categories: supervised and unsupervised. In supervised learning, models are trained on well-labelled data sets and then tested against the same. This is best suited to regression or classification problems. ‘In the predictive or supervised learning approach, the goal is to learn a mapping from inputs x to outputs y, given a labelled set of input-output pairs’.

Unsupervised learning is useful when perfectly labelled data is not available to the machine. A deep learning model is used in conjunction with a dataset without predefined labels for what the machine is supposed to accomplish:

‘Here we are only given inputs, and the goal is to find ‘interesting patterns’ in the data. […] This is a much less well-defined problem, since we are not told what kinds of patterns to look for, and there is no obvious error metric to use (unlike supervised learning, where we can compare our prediction of y for a given x to the observed value)’.

GANs change the definition given above just slightly by using both supervised and unsupervised components. Instead of relying upon training sets of known data, the network is based on a game theory scenario where two neural networks compete with one another. The first neural network, a generator, will compete against a discriminator. The generator is responsible for creating samples (unsupervised) to intertwine within the training set (supervised). The discriminator is responsible for determining if the example it is presented with is from the training set of real data or if it has been given a sample from the generator. This output of this network is a sample that has been created by the generator and which passes the test of the discriminator. The generator attempts to fool the classifier into believing its samples are real, indistinguishable from the real learning data.

This approach is best explained using Goodfellow’s counterfeiter analogy:

The generative model can be thought of as analogous to a team of counterfeiters, trying to produce fake currency and use it without detection, while the discriminative model is analogous to the police, trying to detect counterfeit currency. Competition in this game drives both teams to improve their methods until the counterfeits are indistinguishable from the genuine articles.

GANs work well with image synthesis. The literature suggests that they are most suited to deep fakes which is a type of multimedia where the file is manipulated to appear as real as possible. This is not the only application of GAN technology, however. Other applications have been postulated to include image-to-image translation, text-to-image translation, artificial face aging and enhanced super-resolution. Most often, applications of GAN are within the multimedia space as image and video data is well suited to deep learning methods. Applications within the intelligence space use not just GAN but other machine learning algorithms. In this report, we consider machine learning applications within the cyber intelligence domain, but restrict our assessments to those which use GAN. Further work should be conducted to expand this study into cyber intelligence applications using any machine learning method, not just GAN.

Complete Paper: Generative Adversarial Networks from a Cyber Intelligence Perspective (PDF) – Mouseover to Scroll

Read the original paper.

Reference: Biondi, F., Buonocore, G. and Matthews, R., 2021. Generative Adversarial Networks from a Cyber Intelligence perspective. [online] Tallinn: NATO CCDCOE. Available at: <https://ccdcoe.org/uploads/2021/08/Generative-Adversarial-Networks-from-a-Cyber-Intelligence-view.pdf> [Accessed 25 August 2021].

*Shared with permission as an educational paper for non-commercial use in accordance with NATO CCDCOE disclaimer terms.

Additional Reading

• An eDiscovery Market Size Mashup: 2020-2025 Worldwide Software and Services Overview
• Defining Cyber Discovery? A Definition and Framework

Source: ComplexDiscovery