Storj

A Peer-to-Peer Cloud Storage Network

Shawn Wilkinson ([email protected]),

Tome Boshevski ([email protected]),
Josh Brandoff ([email protected]),
James Prestwich ([email protected]),
Gordon Hall ([email protected]),
Patrick Gerbes ([email protected])
Philip Hutchins ([email protected])
Chris Pollard ([email protected])

With contributions from: Vitalik Buterin ([email protected])

December 15, 2016

v2.0

Abstract

A peer-to-peer cloud storage network implementing client-side encryp-

tion would allow users to transfer and share data without reliance on a
third party storage provider. The removal of central controls would mit-
igate most traditional data failures and outages, as well as significantly
increase security, privacy, and data control. Peer-to-peer networks are
generally unfeasible for production storage systems, as data availability
is a function of popularity, rather than utility. We propose a solution in
the form of a challenge-response verification system coupled with direct
payments. In this way we can periodically check data integrity, and of-
fer rewards to peers maintaining data. We further propose a model for
addressing access and performance concerns with a set of independent or
federated nodes.

1
1 Introduction

Cloud storage has come to rely almost exclusively on large storage providers act-
ing as trusted third parties to transfer and store data. This system suffers from
the inherent weaknesses of a trust-based model. Because client-side encryp-
tion is nonstandard, the traditional cloud is vulnerable to a variety of security
threats, including man-in-the-middle attacks, malware, and application flaws
that expose private consumer and corporate data. Moreover, because many
storage devices rely on the same infrastructure, failures is correlated across files
and systems.

A decentralized cloud storage network offers many advantages compared to

datacenter-based cloud storage. Data security can be maintained using client-
side encryption, while data integrity will be maintained via a proof of retrievabil-
ity. The impact of infrastructure failures and security breaches will be greatly
reduced. An open market for data storage may drive down costs for various
storage services by enabling more parties to compete using existing devices.
Data on the network will be resistant to censorship, tampering, unauthorized
access, and data failures. This paper describes a concrete implementation of
such a network, and a set of tools for interacting with that network.

2 Design

Storj is a protocol that creates a distributed network for the formation and
execution of storage contracts between peers. The Storj protocol enables peers
on the network to negotiate contracts, transfer data, verify the integrity and
availability of remote data, retrieve data, and pay other nodes. Each peer is
an autonomous agent, capable of performing these actions without significant
human interaction. Many of the basic tools for these interactions are described
in this document. Full protocol documentation can be found elsewhere [1].

2.1 Files as Encrypted Shards

A shard is a portion of an encrypted file to be stored on this network. Sharding

has a number of advantages to security, privacy, performance, and availability.

Files should be encrypted client-side before being sharded. The reference

implementation uses AES256-CTR, but convergent encryption or any other de-
sirable system could be implemented. This protects the content of the data

2
from the storage provider, or farmer, housing the data. The data owner retains
complete control over the encryption key, and thus over access to the data.

The data owner may separately secure knowledge of how a file is sharded
and where in the network the shards are located. As the set of shards in the
network grows, it becomes exponentially more difficult to locate any given shard
set without prior knowledge of their locations (see Section 6.3). This implies
that security of the file is proportional to the square of the size of the network.

Shard size is a negotiable contract parameter. To preserve privacy, it is

recommended that shard sizes be standardized as a byte multiple, such as 8 or
32 MB. Smaller files may be filled with zeroes or random data. Standardized
sizes dissuade side-channel attempts to determine the content of a given shard,
and can mask the flow of shards through the network.

Sharding large files like video content and distributing the shards across
nodes reduces the impact of content delivery on any given node. Bandwidth
demands are distributed more evenly across the network. In addition, the end-
user can take advantage of parallel transfer, similar to BitTorrent [2] or other
peer-to-peer networks.

Because peers generally rely on separate hardware and infrastructure, data

failure is not correlated. This implies that creating redundant mirrors of shards,
or applying a parity scheme across the set of shards is an extremely effective
method of securing availability. Availability is proportional to the number of
nodes storing the data.

Figure 1: Visualizing the Sharding Process

1. Files are encrypted.

2. Encrypted files are split into shards, or multiple files are combined to form
a shard.

3
3. Audit pre-processing is performed for each shard (see Section 2.3).

4. Shards may be transmitted to the network.

2.2 Kademlia and Modifications

Storj is built on Kademlia [3], a distributed hash table (DHT). It is important

to note that shards are not stored in the hash table. Rather, Kademlia creates a
distributed network with efficient message routing and other desirable qualities.
Storj adds several message types, and enhancements to core Kademlia function-
ality (see Appendix A). In the future, the hash table may be used as a store for
data location information, or other purposes.

2.2.1 Signature Verification

Similar to S/Kademlia [4], the Storj network requires peers to sign messages.
To join the network a node must create an ECDSA keypair, (kpriv , kpub ). The
Kademlia Node ID corresponds to ripemd160(sha256(kpub )). As such, each
Node ID in the Storj network is also a valid Bitcoin address, which the node
can spend from. Nodes sign all messages, and validate message signatures be-
fore processing messages. This modification enforces long-term identity on the
network, and provides a proof of work deterrent to Eclipse attacks on Kademlia
routing (see Section 5.2). In the future there are a variety of other uses for this
address.

2.3 Proofs of Retrievability

Proofs of retrievability guarantee the existence of a certain piece of data on a

remote host. The ideal proof minimizes message size, can be calculated quickly,
requires minimal pre-processing, and provides a high degree of confidence that
the file is available and intact. To provide knowledge of data integrity and
availability to the data owner, Storj provides a standard format for issuing and
verifying proofs of retrievability via a challenge-response interaction called an
audit or heartbeat.

Our reference implementation uses Merkle trees [5] and Merkle proofs. After
the sharding process the data owner generates a set of n random challenge salts
s0 , s1 , ...sn−1 and stores the set of salts s. The challenge salts are each prepended
to the data d, and the resulting string is hashed to form a pre-leaf p as such:
pi = H(si + d). Pre-leaves are hashed again, and the resulting digests become

4
the set of leaves l of a standard Merkle tree such that li = H(H(si + d)). The
leaf set is filled with hashes of a blank string until its cardinality is a power of
two, to simplify the proof process.

Figure 2: Storj Audit Tree with |l| = 4

Red outlines indicate the elements of a Merkle proof for s0

The data owner stores the set of challenges, the Merkle root and the depth
of the Merkle tree, then transmits the Merkle trees leaves to the farmer. The
farmer stores the leaves along with the shard. Periodically, the data owner
selects a challenge from the stored set, and transmits it to the farmer. Challenges
may be selected according to any reasonable pattern, but should not be reused.
The farmer uses the challenge and the data to generate the pre-leaf. The pre-
leaf, along with the set of leaves, is used to generate a Merkle proof, which is
sent back to the data owner.

The Storj Merkle proof always consists of exactly log2 (|l|) + 1 hashes, and
thus is a compact transmission, even for large trees. The data owner uses
the stored Merkle root and tree depth to verify the proof by verifying that its
length is equal to the tree depth and the hashes provided recreate the stored
root. This scheme does not allow false negatives or false positives, as the hash
function requires each bit to remain intact to produce the same output.

5
2.3.1 Partial Audits

The Merkle tree audit scheme requires significant computational overhead for
the data owner, as the entire shard must be hashed many times to generate
pre-leaves. An extension of this scheme utilizes subsets of the data to perform
partial audits, reducing computational overhead. This also has the advantage
of significantly reducing I/O burden on farmer resources.

This extension relies on two additional selectable parameters: a set of byte

indices x within the shard and a set of section lengths in bytes, b. The data
owner stores a set of 3-tuples (s, x, b). To generate pre-leaf i, the data owner
prepends si to the bi bytes found at xi . During the audit process, the verifier
transmits (s, x, b)i , which the farmer uses to generate a pre-leaf. The Merkle
proof is generated and verified as normal.

Figure 3: Storj Audit Tree with |l| = 4 and Partial Audits

Red outlines indicate the elements of a Merkle proof for s0

Partial audits provide only probabilistic assurance that the farmer retains
the entire file. They allow for false positive results, where the verifier believes the
farmer retains the intact shard, when it has actually been modified or partially
deleted. The probability of a false positive on an individual partial audit is
easily calculable (see Section 6.4)

6
Thus the data owner can have a known confidence level that a shard is
still intact and available. In practice, this is more complex, as farmers may
implement intelligent strategies to attempt to defeat partial audits. Fortunately,
this is a bounded problem in the case of iterative audits. The probability of
several consecutive false positives becomes very low, even when small portions
of the file have been deleted.

In addition, partial audits can be easily mixed with full audits without re-
structuring the Merkle tree or modifying the proof verification process. Many
audit strategies that mix full and partial verification can be envisioned, each of
which provides different levels of confidence over time.

A further extension of this scheme could use a deterministic seed instead of

a set of byte indexes. This seed would be used to generate indexes of many non-
consecutive bytes in the file. Requiring many non-consecutive random bytes
would provide additional resistance against malicious farmers attempting to
implement audit evasion strategies without significant extra overhead from pro-
cessing or I/O.

2.3.2 Other Proof-of-Retrievability Schemes

Other audit schemes were examined, but deemed generally unfeasible. For ex-
ample, Shacham and Waters proposed a compact proof [6] with several advan-
tages over Merkle-tree schemes. This construction allows for an endless stream
of challenges to be generated by the data owner with minimal stored informa-
tion. It also allows for public verifiability of challenge responses.

However, initial implementations indicate that the client-side pre-processing

required for the Shacham-Waters scheme requires at least one order of magni-
tude more computation time than hash-based methods, rendering it too slow
for most applications.

Proof of retrievability is an area of ongoing research, and other practical

schemes may be discovered in the future. As proof of retrievability schemes
are discovered and implemented, the choice of scheme may become a negotiable
contract parameter. This would allow each data owner and node to implement
a wide variety of schemes, and select the most advantageous scheme for a given
purpose.

7
2.3.3 Issuing Audits

To issue audits, Storj extends the Kademlia message set with a new type: AU-
DIT (for a full list of Kademlia extensions, see Appendix A). These messages
are sent from data owners to farmers and contain the hash of the data and a
challenge. The farmer must respond with a Merkle proof as described above.
Upon receipt and validation of the Merkle proof, the data owner must issue
payment to the farmer according to agreed-upon terms.

Figure 4: Issuing and Verifying Storj Audits

2.4 Contracts and Negotiation

Data storage is negotiated via a standard contract format [7]. The contract is a
versioned data structure that describes the relationship between data owner and
farmer. Contracts should contain all information necessary for each node to form
a relationship, transfer the data, create and respond to audits over time, and
arbitrate payments. This includes shard hash, shard size, audit strategy, and
payment information. Storj implements a publish/subscribe system to connect
parties interested in forming a contract (see Section 2.6).

Each party should store a signed copy of the contract. Contracts exist solely
for the benefit of the data owner and farmer, as no other node can verify the
terms or state of the relationship. In the future, contract information may be
stored in the DHT, or in an external ledger like a Blockchain, which may allow
some outside verification of relationship terms.

The contracting system extends Kademlia with four new message types:
OFFER, CONSIGN, MIRROR, and RETRIEVE.

To negotiate a contract, a node creates an OFFER message and sends it

to a prospective partner. Prospective partners are found via the publish/sub-

8
scribe system described in Section 2.6. The OFFER message contains a fully-
constructed contract that describes the desired relationship. The two nodes
repeatedly swap signed OFFER messages. For each new message in the OF-
FER loop, the node either chooses to terminate negotiations, respond with a
new signed counter-offer, or accept the contract by countersigning it. Once an
OFFER is signed by both parties, they each store it locally, keyed by the hash
of the data.

Once an agreement is reached, the data owner sends a CONSIGN message

to the farmer. The message contains the leaves of the audit-tree. The farmer
must respond with a PUSH token that authorizes the data owner to upload the
data via HTTP transfer (see Section 2.10). This token is a random number,
and can be delegated to a third party.

RETRIEVE messages signify the intent to retrieve a shard from a farmer.

These messages are nearly identical to CONSIGN messages, but do not contain
the audit-tree leaves. The farmer responds to a valid RETRIEVE message
with a PULL token that authorizes download of the data via a separate HTTP
transfer.

MIRROR messages instruct a farmer to retrieve data from another farmer.

This allows data owners to create redundant copies of a shard without expending
significant additional bandwidth or time. After a successful OFFER/CONSIGN
process, the data owner may initiate a separate OFFER loop for the same data
with another farmer. Instead of issuing a CONSIGN message to the mirroring
farmer, the data owner instead issues a RETRIEVE message to the original
farmer, and then includes the retrieval token in a MIRROR message to the
mirroring farmer. This authorizes the mirroring farmer to retrieve the data
from the original farmer. The success of the MIRROR process should be verified
immediately via an AUDIT message.

2.5 Payment

Storj is payment agnostic. Neither the protocol nor the contract requires a
specific payment system. The current implementation assumes Storjcoin, but
many other payment types could be implemented, including BTC, Ether, ACH
transfer, or physical transfer of live goats.

The reference implementation will use Storjcoin micropayment channels,

which are currently under development [8]. Micropayment channels allow for
pairing of payment directly to audit, thus minimizing the amount of trust nec-
essary between farmers and data owners. However, because data storage is

9
inexpensive, audit payments are incredibly small, often below $0.000001 per
audit.

Storjcoin allows much more granular payments than other candidate curren-
cies, thereby minimizing trust between parties. In addition, the mechanics of
micropayment channels require the total value of the channel to be escrowed for
the life of the channel. This decreases currency velocity, and implies that value
fluctuations severely impact the economic incentives of micropayment channels.
The use of a separate token creates a certain amount of insulation from outside
volatility, and Storjcoin’s large supply minimizes the impact of token escrow on
the market.

New payment strategies must include a currency, a price for the storage, a
price for retrieval, and a payment destination. It is strongly advised that new
payment strategies consider how data owners prove payment, and farmers verify
receipt without human interaction. Micropayment networks, like the Lightning
Network [9], Implementation details of other payment strategies are left as an
exercise for interested parties.

2.6 Quasar

Storj implements a peer-to-peer publish/subscribe system called Quasar [10][11].

Quasar offers topic-based pub/sub utilizing Bloom filters [12]. Topics describe
ranges of contract parameters the broadcasting node desires, including contract
size and bandwidth commitment. The topic list is standardized on the pro-
tocol level, and easily extensible [13]. This facilitates the contract offer and
negotiation process by ensuring messages reach interested nodes (see Section
2.4)

To operate Quasar, Storj extends Kademlia with three new message types:
SUBSCRIBE, UPDATE, and PUBLISH. These messages facilitate the creation
and propagation of filters. Each node maintains information about topics to
which it subscribes, as well as topics to which its neighbors subscribe in the
form of an attenuated Bloom filter with depth K = 3. The filter at level 1
represents the subscriptions of nodes one hop away.

SUBSCRIBE requests filter lists from neighbors. To build an initial filter

list, nodes issue SUBSCRIBE messages to their three nearest neighbors. Nodes
respond to a SUBSCRIBE with their current filter list. The requesting node
adds the received filter lists to its own attenuate Bloom filter.

UPDATE pushes local filter changes to the three nearest neighbors. When

10
a node subscribes to a new topic it must issue a SUBSCRIBE request to learn
about its neighbors states, then an UPDATE request to notify neighbors of its
new subscription. By exchanging filter lists via SUBSCRIBE/UPDATE loops
nodes gain progressively better knowledge of what information is desirable to
their neighbors, and to nodes reachable by their neighbors.

PUBLISH broadcasts a message to the network. In Storj, these are typi-

cally public announcements of partially-constructed contracts (see Section 2.4).
PUBLISH messages are sent to a nodes three nearest neighbors, and include a
topic parameter whose digest is compared to each filter in the filter list. If the
topic digest is found in the filter at K = 0, the node processes the message. If
the topic is found in any other filters in the filter list, the node forwards the
message to its neighbors. If no matching filter is found, the message is forwarded
to a randomly selected peer in that nodes routing table.

To prevent re-circulation of messages, nodes add their node ID to PUBLISH

messages when they are forwarded, indicating that that node has already re-
ceived the message. When forwarding messages, nodes ignore other nodes whose
IDs are present in that list. This prevents redundant recirculation of messages
with minimal communication overhead.

PUBLISH messages also include a time-to-live (TTL) parameter, measured

in hops. Nodes will not relay messages whose hops have exceeded the TTL.
To prevent spam attacks, nodes will also refuse to relay messages whose TTL
exceeds the TTL that node would give a new message.

Because of this routing process based on subscription filters, PUBLISH mes-

sages propagate through the network randomly until they find a node whose
filter list contains a subscription or a false positive. Once a subscription is in-
dicated by a filter, the message moves towards the subscriber quickly. If a false
positive is encountered the message resumes random routing.

2.7 Redundancy Schemes

Cloud object stores typically own or lease servers to store their customers files.
They use RAID schemes or a multi-datacenter approach to protect the file from
physical or network failure. Because Storj objects exist in a distributed network
of untrusted peers, farmers should not be relied upon to employ the same safety
measures against data loss as a traditional cloud storage company. Indeed,
farmers may simply turn off their node at any time. As such, it is strongly
recommended that the data owner implement redundancy schemes to ensure the
safety of their file. Because the protocol deals only with contracts for individual

11
shards, many redundancy schemes may be used. Three are described below.

2.7.1 Simple Mirroring

The simplest solution is to mirror shards across several nodes. Mirroring pro-
tects against hardware failures by ensuring that multiple copies of each shard
Qn
exist. Availability of the shard with this scheme is P = 1 − 0 an where an
is the uptime of the node storing shard n. Because all shards are required to
assemble the file, availability of the file is equal to the availability of the least
available shard. In the case of a dropped contract, a redundant copy of that
shard can be retrieved and a new location found for it on the network. This is
the current behavior of the reference implementation.

2.7.2 K-of-M Erasure Coding

Storj will soon implement client-side Reed-Solomon erasure coding [14]. Erasure
coding algorithms break a file into k shards, and programmatically create m
parity shards, giving a total of k + m = n shards. Any k of these n shards can
be used to rebuild the file or any missing shards. Availability of the file is then
Qm
P = 1 − 0 am across the set of the m + 1 least available nodes. In the case
of loss of individual shards, the file can be retrieved, the missing shard rebuilt,
and then a new contract negotiated for the missing shard.

To prevent loss of the file, data owners should set shard loss tolerance levels.
Consider a 20-of-40 erasure coding scheme. A data owner might tolerate the loss
of 5 shards out of 40, knowing that the chance of 16 more becoming inaccessible
in the near future is low. However, at some point the probabilistic availability
will fall below safety thresholds. At that point the data owner must initiate a
retrieve and rebuild process.

Because node uptimes are known via the audit process, tolerance levels may
be optimized based on the characteristics of the nodes involved. Many strategies
may be implemented to handle this process.

Erasure coding is desirable because it drastically decreases the probability of

losing access to a file. It also decreases the on-disk overhead required to achieve
a given level of availability for a file. Rather than being limited by the least
available shard, erasure coding schemes are limited by the least-available n + 1
nodes (see Section 6.1).

12
2.8 KFS

To facilitate on-disk storage for farmers, Storj implements a local file store
called KFS [15]. The farming client initially used the filesystem directly to
store shards. Later the farming client used a single LevelDB instance. Both of
these approaches failed to scale. For example, LevelDB compaction processing
time scales linearly with store size, and locks both reads and writes while in
progress. This significantly impacted performance and availability for nodes
storing more than 100GB. KFS is an abstraction layer over a set of LevelDB
instances that seeks to address scaling problems.

2.8.1 Rationale

LevelDB is a key-value store. It has many desirable qualities, including long-

term support, portability, and high performing reads powered by lexicographi-
cally sorted keys. While LevelDB compaction is typically a desirable feature, it
severely limits scaling. Its impact is larger on lower end systems and can also
vary based on the type of disk in use. Compaction also blocks reads and writes
during this period, rendering Storj nodes effectively offline until the process
completes.

However, because LevelDB instances are cheap to create, open, and close,
compaction costs can be bounded by managing a set of size-limited LevelDB
instances. Instances can be initialized ad hoc, and opened and closed as neces-
sary.

Horizontally scaling many LevelDB instances has a number of benefits to

scalability. Chiefly, it mitigates the impact compaction has on operations. Be-
cause compaction runs individually across each instance, rather than across the
whole data set, the issues compaction causes for scaling are minimized. Al-
though compaction across the shard set as a whole will take approximately the
same amount of computation (compaction scales linearly with data), it now oc-
curs separately for each instance. Which is to say, compaction is broken up into
256 smaller processes running independently.

With KFS compaction locks only individual buckets, leaving many others
available to read and write. Operations and compaction are distributed evenly
across hundreds of buckets, meaning the chance of an operation being blocked
by compaction is small. Whereas previously compaction would block the entire
shard set for several seconds (or longer on low-end hardware), it now blocks
only small sections of the shard set for a much shorter time period.

13
2.8.2 S-Buckets and Routing

Rather than a single large instance, KFS stores shards in B size-limited LevelDB
instances, called S-Buckets. Collectively S-Buckets L0 , L1 , LB−1 form the B-
Table. S-Buckets have a fixed maximum size in bytes, S. Thus the maximum
size of a KFS store is S ∗ B bytes. Storj currently uses S = 32 GiB and B = 256
for a total capacity of 8 TiB.

KFS requires that there be a reference identifier, which can be any arbitrary
R bit key where R ≥ log2 (B). Storj nodes will use their 160 bit Node ID.
Incoming shards are sorted into S-bucket according to the following method:

1. Let g = dlog2 (B)e.

2. Let h be the first g bits of R.

3. Let i be the first g bits of the shard hash.

4. Let n = h ⊕ i.

5. Store the shard in Ln .

This sorting algorithm is fast, deterministic, and uses only readily-available

information. Where B is a power of two, it also provides an even distribution
of shards across all S-Buckets, as shown in Figure 5 below.

Figure 5: XOR Distance of Random Bytes

S-Buckets that have reached S bytes cannot store more shards. Farmers
can determine if bucket Ln is full during contract negotiation by calculating n

14
using the data-hash field in the contract, and should reject contracts that would
cause them to overfill an S-Bucket. When all buckets in a KFS instance are
full, a second instance may be made. Given that the 8 TiB upper limit of a
farmers KFS instance is larger than most available drives, this is unlikely for
most hardware.

2.8.3 Keying By Shard Hash

As mentioned earlier, LevelDB sorts items lexicographically by key. KFS takes

advantage of this to optimize the efficiency of reads and writes. Data is stored
in chunks of C bytes (or less). By default C = 128 KiB. These chunks are keyed
by the full content’s hash followed by a space and a numerical index. This
ensures that key/value pairs are small, and that reads and writes to and from
a S-Bucket are sequential. It also allows for efficient streaming of data both in
and out of the S-bucket.

Because of the idiosyncrasies of sorting numerical strings lexicographically,

the index substring should be expressed in base 10, and have a constant number
S
of characters. The length l of the index string should be defined by l = dlog10 C e.
For default parameters l = 6. Therefore the chunk at i = 3753 will have
the index number ’003753’. This ensures that chunks of a shard are stored
consecutively.

To preserve order and therefore maximize read/write performance, the keys

for the chunks of a specific shard should be strings generated as follows:

1. Determine the chunk index number, and encode it as a string.

2. Prepend the chunk index string with ’0’ until it reaches length l.

3. Encode H(data) as a string of hexadecimal characters.

4. Append a single space to the hash.

5. Append the modfied chunk index to the hash.

2.8.4 Performance Benefits

In initial testing KFS outperforms vanilla LevelDB in reads, writes, and unlinks
for a variety of file sizes. KFS displays lower means and lower variance across
almost all combinations of file size, operation, and storage device type. It par-
ticularly excels at unlinks and writes of large files, reducing variance by several

15
orders of magnitude. The full methodology of these tests, and their results, can
be found elsewhere [16].

Figure 6: Relative Performance of KFS and LevelDB

2.9 NAT Traversal and Reverse HTTP Tunneling

Due to the presence of NATs and other adverse network conditions, not all
devices are publicly accessible. To enable non-public nodes to participate in the
network, Storj implements a reverse tunnel system.

To facilitate this system, Storj extends Kademlia with three additional mes-
sage types: PROBE, FIND TUNNEL, and OPEN TUNNEL The tunneling sys-
tem also makes use of the publish/subscribe system detailed in section 2.6.

PROBE messages allow a node to determine whether it is publically address-

able. The message is sent to a publicly addressable node, typically a known net-
work seed. The receiving node issues a separate PING message. The receiving
node then responds to the PROBE message with the result of the PING. Nodes
joining the network should immediately send a PROBE to any known node.

Nodes that receive a negative response to their initial PROBE should issue
a FIND TUNNEL request to any known node. That node must respond with
three contacts that have previously published a tunnel announcement via the
publish/subscribe system. Tunnel providers must be publicly addressable.

16
Once the non-public node has received a list of tunnel providers, it issues
OPEN TUNNEL requests to the tunnel providers. The providers must provide
a tunnel for that node if they are capable. To open a connection, the provider
sends back an affirmative response with tunnel information. The tunneled node
then opens a long-lived connection to the provider, and updates its own contact
information to reflect the tunnel address.

Tunnels are operated over TCP sockets by a custom reverse-tunneling li-

brary, Diglet [17]. Diglet provides a simple and flexible interface for general-
purpose reverse tunneling. It is accessible both by command-line and program-
matically.

2.10 Data Transfer

Data is transferred via HTTP [18]. Farmers expose endpoints where client
applications may upload or download shards. Clients’ requests are authenticated
via tokens provide by previous CONSIGN and RETRIEVE messages. This
transfer mechanism is not essential to the protocol, and many alternatives may
be implemented in the future.

3 Network Access

As should be apparent, the data owner has to shoulder significant burdens to

maintain availability and integrity of data on the Storj network. Because nodes
cannot be trusted, and hidden information like challenge sets cannot be safely
outsourced to an untrusted peer, data owners are responsible for negotiating con-
tracts, pre-processing shards, issuing and verifying audits, providing payments,
managing file state via the collection of shards, managing file encryption keys,
etc. Many of these functions require high uptime and significant infrastructure,
especially for an active set of files. User run applications, like a file syncing
application, cannot be expected to efficiently manage files on the network.

To enable simple access to the network from the widest possible array of
client applications, Storj implements a thin-client model that delegates trust to
a dedicated server that manages data ownership. This is similar to the SPV
wallet concept found in Bitcoin and other cryptocurrency ecosystems. The
burdens of the data owner can be split across the client and the server in a
variety of ways. By varying the amount of trust delegated, the server could also
provide a wide variety of other valuable services. This sort of dedicated server,

17
called Bridge, has been developed and released as Free Software. Any individual
or organization can run their own Bridge server to facilitate network access.

3.1 Bridge

Our reference implementation of this model consists of a Bridge server, and a

client library. Bridge provides an object store, which is to say, the primary
function of Bridge is to expose an API to application developers. Developers
should be able to use the Bridge via a simple client without requiring knowl-
edge of the network, audit procedures, or cryptocurrencies. The Bridge API
is an abstraction layer that streamlines the development process. This enables
developers to create many applications that use the Storj network, allowing the
network to reach many users.

In the current implementation, Bridge assumes responsibility for contract

negotiation, audit issuance and verification, payments, and file state, while the
client is responsible for encryption, pre-processing, and file key management.
The Bridge exposes access to these services through a RESTful API. In this
way, the client can be completely naive of the Storj protocol and network while
still taking advantage of the network. In addition, because the dedicated server
can be relied on to have high uptime, the client can be integrated into unreliable
user-space applications.

Bridge is designed to store only metadata. It does not cache encrypted shards
and, with the exception of public buckets, does not hold encryption keys. The
only knowledge of the file that Bridge is able to share with third parties is
metadata such as access patterns. This system protects the client’s privacy and
gives the client complete control over access to the data, while delegating the
responsibility of keeping files available on the network to Bridge.

It is possible to envision Bridge upgrades that allow for different levels of

delegated trust. A Bridge client may want to retain control over issuing and
validating audits, or managing pointers to shards. Or a client may choose
to authorize two or more unrelated Bridges to manage its audits in order to
minimize the trust it places in either Bridge server. In the long run, any function
of the data owner can be split across two or more parties by delegating trust.

18
3.2 Bridge API and Client

Full documentation of the Bridge API is outside the scope of this whitepaper,
but is available elsewhere [19]. The first complete client implementation is in
JavaScript. Implementations in C, Python, and Java are in progress.

Because files cannot simply be POSTed to API endpoints, the structures

of the Bridge API and client are different from existing object stores. Clients
are implemented to hide the complexity of managing files on the storj network
through simple and familiar interfaces. As much as possible, complex network
operations are abstracted away behind simple function calls.

A brief summary of the upload process follows:

1. The client gathers and pre-processes data.

2. The client notifies Bridge of data awaiting upload.

3. Bridge negotiates contracts with network nodes.

4. Bridge returns the IP addresses of contracted nodes, and authorization

tokens to the client.

5. The client uses the IP addresses and tokens to contact the farming nodes
and upload the data.

6. The client transfers the audit information to the Bridge, delegating trust.

7. Bridge immediately issues an audit and verifies the response, to prove data
was transferred correctly.

8. Bridge assumes responsibility for issuing audits, paying farmers, and man-
aging file state.

9. Bridge exposes file metadata to the client via the API.

The download process is similar.

1. The client requests a file by an identifier.

2. Bridge validates the request and provides a list of farmer IP addresses and
tokens.

3. The client uses the addresses and tokens to retrieve the file

4. The file is reassembled and decrypted client-side.

19
The JavaScript library accepts file, handles pre-processing, and manages
connections as directed by Bridge. It also makes decrypted downloads available
to applications as files, or as streams. A sample CLI using the library is available
as Free Software at https://github.com/storj/core-cli. It has been tested with a
wide variety of file sizes, and is capable of reliably streaming 1080p video from
the Storj network.

3.2.1 Application Development Tools

The primary function of Bridge and the Bridge API is to serve applications. To
this end clients and tools in a wide variety of languages are under development.

Storj.js[20] seeks to provide a standard in-browser interface for downloading

files from Storj. Though in early stages, it can already communicate with Bridge,
retrieve file pointers and tokens, retrieve shards from farmers, reassemble shards,
and append the completed file to the DOM. This allows web developers to easily
reference Storj objects from within a page, and rely on them being delivered
properly to the end user. This could be used to provide any service from in-
browser document editing to photo storage.

Key and file management tools for web backends are in early planning stages,
including Storj plugins for standard backend tools like content management
systems. These tools should help content-driven application developers work
with files on the Storj network. Standardizing these tools around permissioning
files by user could help create data portability between services as discussed in
section 4.2.

Bridges to other protocols and workflows are also planned. The Storj CLI
lends itself to shell scripting automation. Similar tools for FTP, FUSE, and
common tools for interacting with files will be developed in the future.

3.3 Bridge as an Authorization Mechanism

Bridge can be used to manage authorization for private files stored on the net-
work. Because Bridge manages the state of each contract under its care, it is
a logical provider of these services. It can manage a variety of authorization-
related services to enable sharing and collaboration.

20
3.3.1 Identity and Permissioning

The Bridge API uses public-key cryptography to verify clients. Rather than the
Bridge server issuing an API key to each user, users register public keys with
the Bridge. API requests are signed, and the Bridge verifies that the signature
matches a registered public key. Bridge organizes file metadata into buckets to
facilitate management. Buckets can be permissioned individually by registering
a set of public keys to the Bucket.

Application developers can use this to easily delegate permissions to ap-

plications, servers, or other developers. For instance, the developer of a file
syncing service could create a keypair for each user of that service, and divide
each user into a separate Bucket accessible only by that users keypair. Usage of
each Bucket is tracked separately, so users who have exceeded their allotment
could have write permissions revoked programmatically. This provides a logical
separation of user permissions, as well as a variety of organizational tools.

3.3.2 Key Migration

Because shard encryption keys are stored on the device that generated them,
data portability is an issue. The reference implementation of Bridge and the
client facilitate the transfer of file encryption keys between clients in a safe
way. Clients generate a cryptographically strong seed, by default a randomly
generated twelve word phrase. To encrypt a given file, the client generates a
key deterministically based on the seed, Bucket ID and File ID.

The user can import the seed one time to each new device, which per-
manantly keeps the devices syncronized. This also facilitates backup since users
only have to store the seed, not every newly generated file key.

3.3.3 Public Files

Bridge, like other object stores, allows developers to create and disseminate
public files via public Buckets. The Bridge server allows the developer to upload
the encryption key, and then allows anonymous users to retrieve the file key
and the set of file pointers. Public Buckets are useful for content delivery to
webpages, or to public-facing applications.

A system to share and retrieve public files without need of a Bridge could
also be created. Pointers and keys could be posted publicly on any platform, and
clients could be required to pay farmers directly for downloads. In practice this

21
would be very similar to an incentivized torrent. Platforms serving pointers
function similarly to trackers facilitating torrents. It is unclear whether this
system would have significant advantages over existing torrent networks.

3.3.4 File Sharing

In the future, the Bridge could enable sharing of specific files between applica-
tions or users. Because all files co-exist on a shared network, this is a problem
of standardization and identity management.

Bridge could also use a third-party source of identity, like a PGP keyserver
or Keybase[21], to enable secure person-to-person file sharing. A tiered keying
strategy (as used by LastPass[20]) could also allow for the sharing of individual
files. Other cryptographic schemes like proxy re-encryption seem promising.
For a simplified example: if file keys are strongly encrypted and escrowed with
a Bridge, files could be shared to any social media handle that could be au-
thenticated via Keybase. Bridge could send the corresponding client a single
encrypted file key along with a transposition key, thus enabling access to a file
without exposing the file to Bridge, or modifying the file in any way.

A thorough description of these key management schemes is outside the scope

of this paper. It is enough to note that they exist, that many useful strategies
can be implemented in parallel, and that a dedicated Bridge can facilitate them
in many useful ways.

3.4 Bridge as a Network Information Repository

As noted earlier, data owners are responsible for negotiating contracts and man-
aging file state. With enough information about peers on the network, contract
selection becomes a powerful tool for maintaining file state. A Bridge will have
many active contracts with many farmers, and will therefore have access to
information about those farmers. A Bridge could use this information to in-
telligently distribute shards across a set of farmers in order to achieve specific
performance goals.

For instance, via the execution of a contract, a Bridge node gathers data
about the farmers communication latency, audit success rate, audit response
latency, and availability. With minimal additional effort, the Bridge could also
gather information about the nodes available bandwidth. By gathering a large
pool of reliable data about farmers, a Bridge node can intelligently select a set of
farmers that collectively provides a probabilistic guarantee of a certain quality

22
of service.

In other words, the Bridge can leverage its knowledge about peers on the
network to tailor the service to the clients requirements. Rather than a limited
set of service tiers, a Bridge could assemble a package of contracts on the fly to
meet any service requirement. This allows the client to determine the optimal
latency, bandwidth, or location of a file, and have confidence that its goals will
be met. For instance, a streaming video application may specify a need for high
bandwidth, while archival storage needs only high availability. In a sufficiently
large network, any need could be met.

Secure distributed computation is an unsolved problem and, as such, each

Bridge server uses its accumulated knowledge of the network. The Bridge is
able to provide a probabilistic quality of service based on its knowledge the
performance and reliability of farmers that a distributed network alone cannot
provide.

3.5 Bridge as a Service

In cases where the cost of delegating trust is not excessively high, clients may
use third-party Bridges. Because Bridges do not store data and have no access
to keys, this is still a large improvement on the traditional data-center model.
Many of the features Bridge servers provide, like permissioning and intelligent
contracting, leverage considerable network effects. Data sets grow exponentially
more useful as they increse in size, indicating that there are strong economic
incentives to share infrastructure and information in a Bridge.

Applications using object stores delegate significant amounts of trust to the

storage providers. Providers may choose to operate public Bridges as a service.
Application developers then delegate trust to the Bridge, as they would to a
traditional object store, but to a lesser degree. Future updates will allow for
various distributions of responsibilities (and thus levels of trust) between clients
and Bridges. This shifts significant operational burdens from the application
developer to the service-provider. This would also allow developers to pay for
storage with standard payment mechanisms, like credit cards, rather than man-
aging a cryptocurrency wallet. Storj Labs Inc. currently provides this service.

23
4 Future Areas of Research

Storj is a work in progress, and many features are planned for future versions.
There are relatively few examples of functional distributed systems at scale, and
many areas of research are still open.

4.1 Federated Bridges

Bridge nodes could cooperate to share data about the network in a mutually
beneficial federation. This would allow each Bridge to improve the quality of
service that it provides by improving the quality of information available.

Bridges could also, with the consent of users, cooperate to share file metadata
and pointers among themselves. This would allow a user to access their file from
any Bridge, rather than being dependent on a single Bridge. A tiered set of
fallback Bridges storing the same access information is a desirable feature, as it
hedges against downtime from a solo Bridge. Some solvable permissioning issues
may exist, but there is no reason to believe a standard format and algorithm
for syncing state across Bridges may not be developed.

4.2 Data Portability

By encouraging use of data format and access standards, Storj aims to allow
portability of data between applications. Unlike a traditional model, where
control of data is tied to the service used to access the data, data access may
be tied to individual users because Storj forms a common underlying layer.
User data can be tied to persistent cryptographic identities, and authenticated
without exposing data to third parties. Siloing data in applications is a harmful
relic of traditional models. Building cross-compatibility into the future of data
storage greatly improves user privacy and user experience.

Applications implementing these standards would be broadly compatible.

When access is tied to users rather than services, privacy and control are pre-
served. A user may grant access to a service that backs up their hard drive,
which places those files in Storj. The user could separately grant access to a
photo-sharing service, which could then access any photos in the backup. The
user gains seamless portability of data across many applications, and application
developers gain access to a large pool of existing users.

Permissioning in this system may be managed by a service like a Bridge, tied

24
to a web of trust identity via services like Keybase, or handled by a distributed
self-sovereign identity system. Smart contract systems, e.g. Ethereum [22] con-
tracts, seem like a sensible long-term choice, as they can provide file permissions
based on arbitrary code execution. Some problems may exist with respect to
management of the private information required for identity and permissioning
systems, but sufficient solutions likely exist.

While this system represents a significant step up in both usability and value,
there are unmitigable security issues. Unfortunately, as in any cryptographic
system, it is impossible to revoke access to data. Applications may cache data
or forward it to third parties. Users, by definition, trust application developers
to handle their data responsibly. To mitigate these risks, Storj Labs intends
to provide incentives to developers to build free and open-source software. No
application can be completely secure, but auditable code is the best defense of
users privacy and security.

The potential advantages in terms of user experience and privacy are great,
but more research is needed. Many open questions exist with respect to per-
missioning mechanisms. At worst a unified backend powering interoperable
applications provides equivalent security to current data-center based models.
Storj hopes to collaborate with other forward-thinking data-driven projects to
create and advocate for these open standards.

4.3 Reputation Systems

Storj, like many distributed networks, would profit immensely from a distributed
reputation system. A reliable means of determining reputation on a distributed
system is an unsolved problem. Several approaches have been detailed, and some
implemented in practice but none have achieved consensus among researchers
or engineers. A brief review of several of these approaches follows.

One inherent downside of distributing information across a network is the

additional latency required for decisionmaking. It is difficult to say whether any
distributed reputation system can accurately assess the bandwidth, latency, or
availability of peers on a distributed network in a manner suitable to object
storage, especially as market demand for these shifts over time. Nevertheless, a
reliable distributed reputation would be an extremely useful tool for interacting
with and understanding the network.

25
4.3.1 IPFS Bitswap

IPFS presents the concept of BitSwap ledgers [23]. BitSwap ledgers are simple
local accounting of past interactions with other nodes. As IPFS and the BitSwap
protocols are primarily concerned with distribution of objects in a Kademlia-
style DHT, BitSwap ledgers count bytes sent and bytes received. Rather than
attempt to reach global consensus about the reputation state of the system,
these ledgers deal only with one-to-one relationships, and do not account for
latency, bandwidth, availability, or other quality of service factors. Most of
the behavior of nodes is left to the implementer, with some discussion given to
potential exchange strategies. This implies that BitSwap ledgers scale well and
are extremely versatile.

4.3.2 Eigentrust and Eigentrust++

Eigentrust [24] attempts to generalize the ledger approach to generate global

trust values in a distributed system using a transitive-trust model. Nodes keep
and exchange trust vectors. For networks with a large majority of trustworthy
peers, the value of each local trust vector converges to a shared global trust
vector as nodes learn more about the network via information exchange.

Eigentrust++ [25] identifies several attack vectors and modifies Eigentrust to

improve performance and reliability in the presence of malicious nodes. Eigen-
trust++ is currently implemented in NEM [26]. Secure global convergence to a
shared trust value for each node is a key feature for any distributed reputation
system.

4.3.3 TrustDavis

TrustDavis [27] implements reputation as insurance. Nodes provide references

for other nodes in the form of insurance contracts. Nodes seeking a prospective
partner for an economic transaction also seek insurance contracts protecting
them from the actions of that partner. Reputation in this system may be
thought of as a graph, with vertices representing nodes, and directed edges
representing the monetary value that a node is willing to stake on behalf of
another. Nodes that are distant in this graph may still transact by purchasing
a set of insurance contracts that traverses these edges. TrustDavis in practice
thus encounters the same routing problem found on other distributed systems
like the Lightning Network.

26
Denominating trust in terms of monetary value is attractive for an economic
network like Storj, but the mechanics of insurance contracts in a system like this
represent an extremely difficult problem. Notably, because failures to deliver
payment propagate backwards through the insurance route, the financial burden
always falls on the node that trusted an untrustworthy node, rather than the
untrustworthy nodes.

4.3.4 Identity Maintenance Costs

Storj is exploring a reputation system that leverages public Blockchains to solve

a narrow set of identity problems [28]. This system requires nodes to spend
money directly to maintain reputation. Nodes invest in their identity over time
by making small standardized payments to their own Storj network node ID.
Because the ID is a Bitcoin address to which the node holds the private key,
these funds are fully recoupable, except for the miners fees. In this system, nodes
prefer to interact with nodes that have a long history of regular transactions.
Over time these indicate monetary investment in an identity equal to the sum
of the miners fees paid.

The payment required to participate in this system should be significantly

less than the expected return of operating a network node. If set correctly, this
recurring monetary payment for an identity bounds the size and duration of
Sybil attacks without affecting cooperative nodes. Legitimate nodes would eas-
ily recoup their identity expense, while Sybil operators would find their expenses
outstripping their returns. Unfortunately, this approach solves a relatively small
subset of identity issues on the network, and it is difficult to see how it could
be extended to other problem sets.

4.4 OFFER Loop Strategies

Many negotiation strategies can exist and interact via the OFFER loop. Full
exploration of negotiation strategies is beyond the scope of this paper, but a few
interesting areas are immediately apparent. Simple examples include price floors
and ceilings, but complex models could be built to base strategies on market
trends and the subjective value of a shard. Negotiation strategies executed by
autonomous agents are an area of (fascinating) ongoing research. Storj will be
one of the first large-scale machine-driven marketplaces. As such, improving
negotiation efficiency is critical to the long-term efficiency of the market.

27
5 Attacks

As with any distributed system, a variety of attack vectors exist. Many of these
are common to all distributed systems. Some are storage-specific, and will apply
to any distributed storage system.

5.1 Spartacus

Spartacus attacks, or identity hijacking, are possible on Kademlia. Any node

may assume the identity of another node and receive some fraction of messages
intended for that node by simply copying its Node ID. This allows for targeted
attacks against specific nodes and data. This is addressed by implementing
Node IDs as ECDSA public key hashes and requiring messages be signed. A
Spartacus attacker in this system would be unable to generate the corresponding
private key, and thus unable to sign messages and participate in the network.

5.2 Sybil

Sybil attacks involve the creation of large amounts of nodes in an attempt to dis-
rupt network operation by hijacking or dropping messages. Kademlia, because
it relies on message redundancy and a concrete distance metric, is reasonably
resistant to Sybil attacks. A nodes neighbors in the network are selected by
Node ID from an evenly distributed pool, and most messages are sent to at
least three neighbors. If a Sybil attacker controls 50% of the network, it suc-
cessfully isolates only 12.5% of honest nodes. While reliability and performance
will degrade, the network will still be functional until a large portion of the
network consists of colluding Sybil nodes.

5.2.1 Google

The Google attack, or nation-state attack, is a hypothetical variant of the Sybil

attack carried out by an entity with extreme resources. Google attacks are
hard to address, as it is difficult to predict the actions of an organization with
orders of magnitude more resources than the sum of the resources of network
participants. The only reliable defence against a Google attack is to create a
network whose resources are on the same order of magnitude as the attackers.
At that scale, any attack against the network would represent an unsustainable
commitment of resources for such an organization.

28
5.2.2 Honest Geppetto

The Honest Geppetto attack is a storage-specific variant of the Google attack.

The attacker operates a large number of puppet nodes on the network, accu-
mulating trust and contracts over time. Once he reaches a certain threshold
he pulls the strings on each puppet to execute a hostage attack with the data
involved, or simply drops each node from the network. Again, the best defence
against this attack is to create a network of sufficient scale that this attack
is ineffective. In the meantime, this can be partially addressed by relatedness
analysis of nodes. Bayesian inference across downtime, latency and other at-
tributes can be used to assess the likelihood that two nodes are operated by
the same organization, and data owners can and should attempt to distribute
shards across as many unrelated nodes as possible.

5.3 Eclipse

An eclipse attack attempts to isolate a node or set of node in the network graph,
by ensuring that all outbound connections reach malicious nodes. Eclipse at-
tacks can be hard to identify, as malicious nodes can be made to function nor-
mally in most cases, only eclipsing certain important messages or information.
Storj addresses eclipse attacks by using public key hashes as Node IDs. In order
to eclipse any node in the network, the attacker must repeatedly generate key
pairs until it finds three keys whose hashes are closer to the targeted node than
its nearest non-malicious neighbor, and must defend that position against any
new nodes with closer IDs. This is, in essence, a proof-of-work problem whose
difficulty is proportional to the number of nodes in the network.

It follows that the best way to defend against eclipse attacks is to increase
the number of nodes in the network. For large networks it becomes prohibitively
expensive to perform an eclipse attack (see Section 6.2). Furthermore, any node
that suspects it has been eclipsed may trivially generate a new keypair and node
ID, thus restarting the proof-of-work challenge.

5.3.1 Tunnel Eclipse

Because tunneled connections rely on the tunnel provider, it is trivial for a tun-
nel provider to eclipse nodes for which it provides tunneled connections. This
attack cannot affect publicly addressable nodes, so it can be trivially defeated
with proper configuration. This attack can be mitigated by encrypting messages
intended for tunneled nodes, thus removing the malicious tunnel provider’s abil-

29
ity to inspect and censor incoming messages. Like a typical eclipse attack, any
node that suspects it is the victim of a tunnel eclipse can easily generate a new
Node ID, and find a new tunnel.

5.4 Hostage Bytes

The hostage byte attack is a storage-specific attack where malicious farmers

refuse to transfer shards, or portions of shards, in order to extort additional
payments from data owners. Data owners should protect themselves against
hostage byte attacks by storing shards redundantly across several nodes (see
Section 2.7). As long as the client keeps the bounds of its erasure encoding
a secret, the malicious farmer cannot know what the last byte is. Redundant
storage is not a complete solution for this attack, but addresses the vast majority
of practical applications of this attack. Defeating redundancy requires collusion
across multiple malicious nodes, which is difficult to execute in practice.

5.5 Cheating Owner

A data owner may attempt to avoid paying a farmer for data storage by refus-
ing to verify a correct audit. In response the farmer may drop the data-owners
shard. This attack primarily poses a problem for any future distributed reputa-
tion system, as it is difficult for outside observers to verify the claims of either
party. There is no known practical publicly verifiable proof of storage, and no
known scheme for independently verifying that a privately verifiable audit was
issued or answered as claimed. This indicates that a cheating client attack is a
large unsolved problem for any reputation system.

5.6 Faithless Farmer

While the farming software is built to require authentication via signature and
token before serving download requests, it is reasonable to imagine a modifica-
tion of the farming software that will provide shards to any paying requestor.
In a network dominated by faithless farmers, any third-party can aggregate and
inspect arbitrary shards present on the network.

However, even should faithless farmers dominate the network, data privacy is
not significantly compromised. Because the location of the shards that comprise
a given file is held solely by the data owner, it is prohibitively difficult to locate
a target file without compromising the owner (see Section 6.3). Storj is not

30
designed to protect against compromised data owners. In addition, should a
third-party gather all shards, strong client-side encryption protects the contents
of the file from inspection. The pointers and the encryption key may be secured
separately. In the current implementation of Bridge, the pointers and the keys
are held by the Bridge and the client, respectively.

5.7 Defeated Audit Attacks

A typical Merkle proof verification does not require the verifier to know the
depth of the tree. Instead the verifier is expected to have the data being vali-
dated. In the Storj audit tree, if the depth is unknown to the verifier the farmer
may attack the verification process by sending a Merkle proof for any hash in
the tree. This proof still generates the Merkle root, and is thus a valid proof of
some node. But, because the verifier does not hold the data used to generate
the tree, it has no way to verify that the proof is for the specific leaf that cor-
responds to the challenge. The verifier must store some information about the
bottom of the tree, such as the depth of the tree, the set of leaves nodes, or the
set of pre-leaves. Of these, the depth is most compact, and thus preferable.

Using the pre-leaf as an intermediary defeats another attack, where the

farmer simply guesses which leaf corresponds to the current challenge. While
this attack is unlikely to succeed, its trivially defeated by forcing the farmer to
provide the pre-leaf. The farmer cannot know the pre-leaf before the challenge
is issued. Requiring transmission of the pre-leaf also allows the data owner to
proceed through the challenge set linearly instead of being forced to select ran-
domly. This is desireable because it allows the data owner to maintain less state
information per tree.

6 Selected Calculations

The following are several interesting calculations related to the operation of the
network.

6.1 Failure of k-of-n Erasure Coding

The chance of failure of k-of-n erasure coding, assuming probability p every

shard stays online, is calculated as a binomial distribution:

31
k−1
X
i n−i n
Prf ailure (n; k, p) = p (1 − p)
i=0
i
n k p Prf ailure n, k, p
18 6 0.5 4.812e-02
18 6 0.75 3.424e-05
18 6 0.9 5.266e-10
18 6 0.98 6.391e-19
36 12 0.5 1.440e-02
36 12 0.75 2.615e-08
36 12 0.9 1.977e-17
36 12 0.98 1.628e-34

Code:
1 def fac ( n ) : return 1 if n ==0 else n * fac (n -1)
2 def choose (n , k ) : return fac ( n ) / fac ( k ) / fac (n - k )
3 def bin (n ,k , p ) : return choose (n , k ) * p ** k * (1 - p ) ** (n - k )
4 def prob _ fail (n ,k , p ) : return sum ([ bin (n ,i , p ) for i in range (0 , k ) ])

Therefore, with well-chosen k and n, in addition to recovery methods de-

scribed above, the statistical chance of shard or file loss is quite small.

6.2 Difficulty of Eclipsing a Target Node

The probability of eclipsing a targeted node in the a network with k nodes in h

hashes is modeled by a similar binomial distribution:
h−1
X 1 h−i h
Prsuccess (h, k) = k −i (1 − )
i=3
k i
h i Prsuccess h, i
100 100 7.937e-02
100 500 1.120e-03
100 900 2.046e-04
500 100 8.766e-01
500 500 8.012e-02
500 900 1.888e-02
900 100 9.939e-01
900 500 2.693e-01
900 900 8.020e-02

Code:

32
1 def fac ( k ) : return 1 if k ==0 else k * fac (k -1)
2 def choose (h , k ) : return fac ( h ) / fac ( k ) / fac (h - k )
3 def bin (i ,h , k ) : return choose (h , i ) * k ** -i * (1 -(1.0/ k ) ) ** (h - i )
4 def prob _ succ (h , k ) : return sum ([ bin (i ,h , k ) for i in range (3 , h ) ])

6.3 Beach Size

As the number of shards on the network grows, it becomes progressively more

difficult to locate a given file without prior knowledge of the locations of its
shards. This implies that even should all farmers become faithless, file privacy
is largely preserved.

The probability of locating a targeted file consisting of k shards by n random

draws from a network containing N shards is modeled as a hypergeometric
distribution with K = k:
N −k

n−k
P rSuccess (N, k, n) = N

n
N k n Prsuccess N, k, n
100 10 10 5.777e-14
100 10 50 5.934e-04
100 10 90 3.305e-01
100 50 50 9.912e-30
100 50 90 5.493e-04
500 50 200 1.961e-22
500 50 400 7.361e-06
900 10 200 2.457e-07
900 10 400 2.823e-04
900 10 800 3.060e-01
900 50 200 1.072e-35
900 50 400 4.023e-19
900 50 800 2.320e-03

Code:
1 def fac ( k ) : return 1 if k ==0 else k * fac (k -1)
2 def choose (h , k ) : return fac ( h ) / fac ( k ) / fac (h - k )
3 def hyp (N ,k , n ) : return choose (N -k ,n - k ) / float ( choose (N , n ) )
4 def prob _ success (N ,k , n ) : return hyp (N ,k , n )

33
6.4 Partial Audit Confidence Levels

Farmers attempting to game the system may rely on data owners to issue partial
audits. Partial audits allow false positives, where the data appears intact, but
in fact has been modified. Data owners may account for this by ascribing
confidence values to each partial audit, based on the likelihood of a false positive.
Partial audit results then update prior confidence of availability. Data owners
may adjust audit parameters to provide desired confidence levels.

The probability of a false positive on a parital audit of n bytes of an N byte

shard, with K bytes modified adversarially by the farmer is a hypergeometric
distribution with k = 0:
N −K

n
P rf alsepositive (N, K, n) = N

n
N K n Prf alsepositive N, K, n
8192 512 512 1.466e-15
8192 1024 512 1.867e-31
8192 2048 512 3.989e-67
8192 3072 512 1.228e-109
8192 4096 512 2.952e-162

Code:
1 def fac ( k ) : return 1 if k ==0 else k * fac (k -1)
2 def choose (h , k ) : return fac ( h ) / fac ( k ) / fac (h - k )
3 def hyp (N ,K , n ) : return float ( choose (N -K , n ) / choose (N , n )
4 def prob _ false _ pos (N ,K , n ) : return hyp (N ,K , n )

As demonstrated, the chance of false positives on even small partial audits

becomes vanishingly small. Farmers failing audits risk losing payouts from cur-
rent contracts, as well as potential future contracts as a result of failed audits.
Dropping 10% of a shard virtually guarantees a loss greater than 10% of the con-
tract value. Thus it stands to reason that partially deleting shards to increase
perceived storage capcity is not a viable economic strategy.

34
A List of Storj Message Types

A.1 Kademlia

1. PING - Determine whether a node is online.

2. STORE - Store a value in the DHT.

3. FIND NODE - Find a node in the DHT.

4. FIND VALUE - Find a value in the DHT.

A.2 Tunneling

5. PROBE - Determine whether the sender is publicly addressable.

6. FIND TUNNEL - Find a publicly addressable node offering tunnels.

7. OPEN TUNNEL - Open a tunnel with a node offering tunnels.

A.3 Quasar

8. SUBSCRIBE - Request filter lists from neighbors.

9. UPDATE - Notify neighbors of updated filters.

10. PUBLISH - Broadcast a message to interested nodes.

A.4 Contracting

11. OFFER - Propose or finalize a contract.

12. CONSIGN - Request a PUSH token from a farmer.

13. MIRROR - Instruct a farmer to retrieve and mirror data from another
farmer.

14. AUDIT - Issue an audit challenge to a farmer.

15. RETRIEVE - Request a PULL token from a farmer.

35
References
[1] G. Hall. Storj core tutorial: Protocol specification, (2016).
https://storj.github.io/core/tutorial-protocol-spec.html.

[2] B. Cohen. Incentives build robustness in bittorrent, (2003).

http://www.bittorrent.org/bittorrentecon.pdf.

[3] P. Maymounkov, D. Mazieres. Kademlia: A peer-to-peer information

system. based on the xor metric, (2002). https://pdos.csail.mit.edu/
~petar/papers/maymounkov-kademlia-lncs.pdf.

[4] I. Baumgart, S. Mies. S/kademlia: A practicable approach towards secure

key-based routing, (2007).
http://www.tm.uka.de/doc/SKademlia 2007.pdf.

[5] R.C. Merkle. Protocols for public key cryptosystems, (April 1980).
http://www.merkle.com/papers/Protocols.pdf.

[6] H. Shacham, B. Waters. Compact proofs of retrievability, (2008).

https://cseweb.ucsd.edu/~hovav/dist/verstore.pdf.

[7] G. Hall. Storj core class: Contract, (2016).

http://storj.github.io/core/Contract.html.

[8] F. Barkhau. Trustless micropayment channels, (2016).

https://github.com/F483/counterparty-documentation/blob/
micropayments/Developers/micropayments.md.

[9] J. Poon, T. Dryja. The bitcoin lightning network: Scalable off-chain

instant payments, (2016).
https://lightning.network/lightning-network-paper.pdf.

[10] G. Hall. Kad quasar, (2016).

https://github.com/kadtools/kad-quasar.

[11] B. Wong, S. Guha. Quasar: A probabilistic publish-subscribe system for

social networks, (2008). http://research.microsoft.com/en-us/um/
people/saikat/pub/iptps08-quasar.pdf.

[12] B. Bloom. Space/time trade-offs in hash coding with allowable errors,

(1970). http://dmod.eu/deca/ft gateway.cfm.pdf.

[13] G. Hall. Storj core tutorial: Publishing storage contracts, (2016).

http://storj.github.io/core/tutorial-contract-topics.html.

36
[14] J. S. Plank. A tutorial on reed-solomon coding for fault-tolerance in
raid-like systems, (1996).
http://web.eecs.utk.edu/~plank/plank/papers/CS-96-332.pdf.

[15] G. Hall. Kfs, (2016). https://storj.github.io/kfs/.

[16] P. Gerbes. Kfs tutorial: Performance testing the changes, (2016).

https://storj.github.io/kfs/tutorial-performance-testing.html.

[17] G. Hall. Diglet server, (2016). http://diglet.me.

[18] G. Hall. Storj core tutorial: Transferring file shards, (2016).

http://storj.github.io/core/tutorial-data-transfer.html.

[19] G. Hall. Storj core, (2016). http://storj.github.io/core/.

[20] LogMeIn Inc. Lastpass password manager. https://www.lastpass.com.

[21] Keybase Inc. Keybase. https://www.keybase.io.

[22] V. Buterin et al. A next-generation smart contract and decentralized

application platform, (2014).
https://github.com/ethereum/wiki/wiki/White-Paper.

[23] J. Benet. Ipfs - content addressed, versioned, p2p file system, (2014).
https://github.com/ipfs/ipfs/blob/master/papers/ipfs-cap2pfs/
ipfs-p2p-file-system.pdf.

[24] S. Kamvar, M. Schlosser, H. Garcia-Molina. The eigentrust algorithm for

reputation management in p2p networks, (2003).
http://ilpubs.stanford.edu:8090/562/1/2002-56.pdf.

[25] X. Fan, L. Liu, M. Li, Z. Su. Eigentrust++: Attack resilient trust

management, (2012). https://pdfs.semanticscholar.org/c5a6/
c05d833179073e3517be6c7e7ab0c9d59b29.pdf.

[26] M. Takemiya et al. Nem technical reference, (2015).

https://www.nem.io/NEM techRef.pdf.

[27] D. DeFigueiredo, E. Barr. Trustdavis: A non-exploitable online reputation

system, (2005). http://earlbarr.com/publications/trustdavis.pdf.

[28] S. Wilkinson, J. Prestwich. Bounding sybil attacks with identity cost,

(2016). https://github.com/Storj/sips/blob/master/sip-0002.md.

37