# Sybil Survey Overview
A [Sybil attack](https://en.wikipedia.org/wiki/Sybil_attack) is an attack on a network service that attempts to subvert the network's purpose through the creation of large numbers of fake users to generate artificial influence. They are extremely difficult to detect using pure automation (i.e., machine learning) so the Gitcoin FDD team uses a human-in-the-loop semi-supervised[1] workflow.

There are several reasons that the FDD team settled on this particular workflow. For one thing, as has been mentioned, it is extremely difficult to identify Sybil attacks using just machine learning. There's a lack of good data for model training. Another reason is ethical. Because it's so difficult at present to detect Sybil attacks, it's important to have multiple levels of detection so that the system as a whole has internal backups and checks. A false positive could be very damaging to a legitimate user. Similarly, through the use of human evaluators we can build good datasets for future machine learning efforts. Finally, there are legal considerations. Under the [General Data Protection Regulation (GDPR)](https://gdpr-info.eu/art-22-gdpr/), people have the right not to be subjected to a purely automated decision making system. Including human review in the process helps keep the system fair and honest.

The [Sybil survey](/4xexQ2dKRu6IeiM0jK68nQ) is part of this human-in-the-loop process for Sybil detection. Human evaluators--volunteers from the Gitcoin FDD team--examine a random subset of Gitcoin contributor data. They flag potential Sybil participants and include a score for how confident they are in their judgment.

The data a group of Google sheets that contain a number of columns and rows. The rows correspond to individual contributors to the current Gitcoin funding round. The columns correspond to data points about that user, both machine- and human-generated.

The human evaluator's job is to look at the users on their assigned sheet, look at the users' behavior and profile as reflected in the data, and come to a decision about whether or not that user is part of a Sybil attack. The evaluator must also assign a score to their confidence in their decision, and leave a short note explaining why they think that user is a Sybil attacker.

---
1. Machine learning algorithms fall into three broad flavors: supervised, unsupervised, and reinforcement learning. You can read more about these flavors in the [contributor's guide](/crGYG06HRfqo3FsnInhrkg) to machine learning algorithms.