IPFS

Security

We approach security from the point of view of the DHT's Keyspace Density.

The following plots examine the peer distribution within the keyspace, aiding in the identification of potential Sybil and eclipse attacks.

Keyspace Density

In Kademlia, every object indexed by the DHT requires a binary identifier. In the libp2p DHT implementation, peers are identified by the digest of `sha256(peer_id)` and CIDs are identified by the digest of `sha256(cid)`. This Kademlia identifier determines the location of an object within the Kademlia XOR keyspace.

Keyspace Regions Population

Keyspace Regions Population

Description: The plot illustrates the number of peers whose Kademlia identifier matches each prefix for all prefixes of a given size, for a given network crawl. Since the density of keyspace regions follows a Poisson distribution, it is expected to observe some regions that are significantly more populated than others.

How to read the plot: The selected `depth` indicates the prefix size. There are `2^i` distinct prefixes at depth `i`. The dropdown menu allows to navigate to previous versions of this plot.

What to look out for: The green dashed line represents the expected density per region, corresponding to the number of peers matching a prefix. A bar significantly exceeding the expected density suggests that a region of the keyspace might be under an eclipse attack, especially if the value is above the risk threshold (red dashed line).

The plot shows how many peers are included in a particular region of the keyspace. Too many peers within one region indicate a potential issue.

Keyspace Density Distribution

Keyspace Density Distribution

Description: As previously mentioned, the keyspace population follows a Poisson distribution, which can make identifying outliers challenging. The plot below counts the number of regions for each population size and facilitates the identification and analysis of outliers. While it is normal for some regions to have populations above the average, the plot enables us to quantify these deviations.

How to read the plot: The green line represents the expected number of regions for each population size. Note that the Poisson distribution is more evident at greater depths (longer prefix size), while analyzing data at lower depths provides limited insights. It is recommended to read the plot for depths between 9 and 13. The date dropdown menu allows to navigate to previous versions of this plot.

What to look out for: If an isolated bar appears on the far right beyond the risk threshold (red dashed) line, it may indicate a potential eclipse attack, warranting further investigation.

The plot shows the distribution of the PeerIDs across the Poisson curve. Too many PeerIDs outside the curve indicate a potential issue.

Topology Dynamics

Although not directly a security metric, very high node churn can affect the performance of a DHT-based network. For instance, Provider Records might become unavailable faster than republication of those records, if the nodes responsible to hold them go offline. In turn, this will result in some keys not being available in the network.

Peer Churn

Peer Churn

This plot presents the Cumulative Distribution Function (CDF) of peer departure times from the network over a measurement period of one-week. The plot basically shows the percentage of nodes online at a point in time (`t=0`) that will have disconnected after 1 hour, 2 hours, 3 hours, .. up to 24 hours after `t=0`.

X-Axis (Time in Hours): Represents the elapsed time since the reference point (`t=0`), measured in hours.

Y-Axis (Percentage of Offline Peers): Indicates the cumulative percentage of peers that have disconnected from the network by the corresponding time interval. Specifically, it shows the proportion of peers that were from projects.common.charts.chart_ids import ChartID online at `t=0` and have since gone offline.

This visualization aids in understanding peer churn dynamics, which is crucial for optimizing network stability, resource allocation, and improving overall decentralized network performance.

Note that the sum of the CDF is NOT 100%, as it only includes peers that were online for 24h, or less.

CDF of Peer Departure Times

Measurement Tools

Our network monitoring employs advanced crawling and probing tools and techniques to gather comprehensive data about network health, topology, and performance