Generative Adversarial Networks, or GANs as they are often called, are a class of artificial intelligence algorithms that are leveraged for unsupervised learning. They were introduced by Ian Goodfellow and his fellow researchers in 2014. GANs are composed of generative models that produce synthetic data samples, and discriminative models that strive to differentiate the synthetic samples from real-world data.

• Two Model System: GANs function based on a system of two neural networks that compete with each other. This process is the crux of unsupervised machine learning in GANs.
• Data Generation: The key function of GANs is to generate new data from scratch. Unlike other machine learning models, GANs generate completely new data that resembles a given dataset.
• Progressive Training: GANs engage in progressive training, where the generated data becomes progressively closer to the real data as the training process evolves.
• Zero-sum Game: GANs operate on a principle called the minimax game, wherein the losses of one player (the discriminator) reflect the gains of the other (the generator), eventually leading to the generation of high-quality data.
• Leveraging Existing Data: GANs are fundamentally designed to learn the distribution of the given dataset and subsequently generate similar data.

Implementing GANs involves several steps starting with a comprehension of the problem scope and the kind of data that needs to be generated. This is followed by defining the architecture and size of the discriminator and generator models, and then training the model iteratively, which leads to a refined GAN that can generate higher quality data. The whole process requires rigorous monitoring and a sound understanding of GANs.

GANs are employed for numerous tasks due to the benefits they offer, which include:

• Data Augmentation: GANs can be used to augment data when the volume of available data is low. This feature of GANs can be of immense help in fields like healthcare, where accessing volumes of data is often problematic.
• Anonymization and Privacy: Because GANs can generate synthetic data that resembles the real-world data, they can be used to create data that maintains the properties of original data while ensuring privacy.
• Pre-training Models: GANs are useful for pre-training models where labelled data is sparse.
• Visual Prowess: GANs have a strong capacity for generating visually coherent images, which is beneficial in applications like virtual reality, video games, and 3D modelling.

While GANs offer numerous advantages, some potential issues include:

• Training Difficulty: GANs can be hard to train. There is a constant tug-of-war between the discriminator and generator, which, when not balanced properly, can lead to a flawed model.
• Computational Complexity: Due to their dual nature, GANs require intensive computation, thus requiring substantial hardware resources.
• Mode Collapse: This occurs when the generator starts producing limited varieties of samples, thereby undermining the model's ability to generate diverse data samples.
• Ethical Concerns: The ability of GANs to generate realistic data raises several ethical concerns such as the generation of deepfakes or counterfeit currency.

In conclusion, GANs are revolutionary in the world of unsupervised machine learning. Despite some existing challenges, with continuous research and advancement in this field, we can anticipate more robust and reliable versions of GANs that can help address a wide range of application-driven challenges.