How ShivaVault protects a video
No single control stops a determined leaker, so ShivaVault stacks them. Some keep a copy from being made at all; the others make sure that any copy that does get out leads back to the one viewing it came from. Each layer below is said first in plain words, then exactly, with how far it is built today.
This is a demo. Two layers run on stand-ins here. DRM uses ClearKey: the stream is encrypted, but the key is handed to the browser, so it shows how DRM works and does not protect the video yet; a licence provider (Widevine, FairPlay, PlayReady) replaces it. Ethereum anchoring runs on Sepolia, a test network: the proofs are real and can be checked there, but a test network is not permanent; the production network replaces it. Do not upload content that needs real protection.
The layers on a viewing
Applied per viewer, while the video plays
-
Visible watermark
BuiltA name tag on every frame: the viewer sees their own name on the picture, and so does anyone they share it with.
Exactly: The viewer's own text is burned into the picture as each segment is delivered. There is no clean copy to serve: every viewer gets their own.
Against: Casual sharing and screen recording; it tells everyone who sees the copy whose it was.
-
Forensic watermark
BuiltA hidden serial number in every frame. Nobody can see it, cutting the name out does not remove it, and only ShivaVault can read it.
Exactly: An invisible mark in every frame carries an opaque payload derived from the viewing session under the tenant's own key. Only Shiva can read it, and only the tenant's index turns it into a viewer.
Against: Copies with the visible text cropped, blurred or covered.
-
A/B forensic watermark
BuiltEvery few seconds each copy differs a little, in a pattern unique to its viewer. Mixing several copies together still points at the people who mixed them.
Exactly: Every 2 s of video carries variant A or B, following a sequence unique to the viewer. The sequence read from a copy points at its viewer with a stated error bound, and at each viewer who mixed copies together.
Against: Collusion between viewers, and copies too degraded for the per-frame mark.
-
Recipient-bound link
DevelopmentA ticket that only works for the person it was sent to: a code by email, then only on their device. A forwarded link plays nothing.
Exactly: A shared link is an invitation: it plays only after a code sent to the recipient's email address, on the one device that redeems it. That device's key, held by the browser and never exported, signs every request. The same step runs in the customer's own web player through Shiva's recipient SDK.
Against: Forwarded and posted links: in anyone else's hands the link plays nothing.
-
DRM
DevelopmentA locked stream: what travels over the internet is useless without a key that only the viewer's player receives while they watch.
Exactly: The viewing is encrypted (CMAF, CENC cbcs) with its own key, released only through a licence to the browser's decryption module. Shaka Player handles it through EME, in the console or in the customer's own web player, which gets the viewing's licence URL. The licence provider plugs in behind Shiva's licence route; none is connected yet, and ClearKey stands in for it in development and on the demo.
Against: Downloading and ripping the stream: the segments are useless without the licence. Hardware-backed Widevine, FairPlay and PlayReady come with the provider.
-
Ethereum anchoring
BuiltA notary's seal nobody can alter: a timestamp on a public blockchain proves each share existed, unchanged, by a given time, without showing what it was.
Exactly: Every viewing and every job output is sealed into a Merkle root on Ethereum every 10 minutes. No share data goes on chain, only the root, so a proof shows the share existed, unchanged, by a given time. It runs on a local chain in development and on Sepolia, a test network, on the demo.
Against: Disputes over whether a share existed, and when: the proof does not depend on trusting Shiva's database.
-
Access trail
BuiltA record of every viewing: when it was watched, from where and on what. A link passed around shows up as many places at once.
Exactly: Every viewing records which addresses and clients played it, minute by minute, and the case file shows it next to a match.
Against: A viewer who denies the viewing was theirs; shared or stolen links show up as many addresses.
-
Ripping alerts
BuiltAn alarm for copying: someone downloading instead of watching is slowed down and noticed, you are told why, and the viewing can be cut off on the spot.
Exactly: Delivery holds every viewing to three times real time, so a download crawls, however many windows, devices or servers it is split over: they all share one limit. The access trail is assessed as it arrives: a viewing pulled faster than real time, with a download tool, in every rendition at once, in several parts at once from parallel windows, from a second device or from many addresses, or one that keeps hitting delivery's limit, is flagged with the reasons on Viewings, the case file and the dashboard, revoked at once if it asked for that, and reported to the customer's own systems as a signed webhook.
Against: A viewer copying a viewing they are allowed to watch: it is caught while it happens, not only once the copy turns up.
Built: built and running. Development: built and running in development and on the demo, with a stand-in where this page says so; not on a customer's service yet. Not built: designed, not built yet.
Every layer comes with every plan, run by us or in your own cloud: see Plans and pricing.
When a copy leaks
From a suspect file to the viewing it came from
-
1
Upload the copy on the console's Forensic cases page.
A file found online, a screen recording, a ripped stream. It goes straight to the tenant's storage with a write-only grant.
-
2
Shiva reads the forensic mark in its frames.
Blind detection: the original is not needed. The mark found names the viewing session, and so its viewer reference, and only this tenant can read it.
-
3
The A/B sequence confirms it, or speaks when the mark cannot.
The copy is cut into 2 s units and each is read as A or B. The string is compared with every A/B viewing of the tenant. A viewer is named only when the chance of a wrong accusation is at most 1 in 1,000; several viewers who mixed their copies are named together as colluders. A short clip may stay inconclusive: roughly 30 s of video is needed to name one viewer among a dozen.
-
4
The case file holds the evidence.
The viewing's session, viewer reference, watermark text and dates, its access trail (addresses and clients by minute), and the Ethereum proof that the share existed, and when, before the leak was found.
-
5
Act on it.
Revoke every live viewing of that viewer at once from the case file, or one by one on Viewings & audit: delivery stops serving them within seconds. A viewing caught ripping may already be revoked by its alert. The rest (contract, takedown) is the customer's, with the case file as evidence.
How the Ethereum proof works
Every share · one root per batch · checked without trusting Shiva
-
1
Each share becomes a leaf, as it is made.
A viewing commits to its asset, its viewer reference, a derivation of its token and its validity; a job output to the file's SHA-256. The leaf is a salted SHA-256 of that: the salt stays with Shiva, so nobody can confirm a leaf by guessing what it hides.
-
2
The leaves fold into one Merkle root.
Every 10 minutes (every minute on the local chain) the period's leaves, of every tenant, are sealed into a Merkle tree. A tenant's proof never shows another tenant's share, only the hashes on the path.
-
3
Only the root goes on chain.
A zero-value transaction from Shiva's own address carries the root, the batch and the leaf count in its input data. There is no contract, no share data and nothing personal on chain. A batch counts as included after 3 confirmations.
-
4
Every share keeps its proof.
Viewings & audit and the case file show it, from pending to sealed, submitted and confirmed: the leaf and its index, the path to the root, the batch, the transaction and its block.
-
5
Anyone can check it, without Shiva.
Rebuild the leaf from the case file's fields and salt, fold it with the path, and compare the result with the root in the transaction on chain. It proves the share existed, unchanged, no later than that block. Who leaked a copy is what the forensic marks tell; the chain proves the share they point at.
AI-assisted detection
Finding a leak sooner · explaining it in plain words
None of this runs yet. It is what Shiva is building next, on top of the marks and the access trail above. The rule that will not change: AI finds and explains, it never decides. Who a copy is traced to, the error bound, a revocation and the Ethereum proof stay with the exact methods described on this page.
-
1
An alert that explains itself.
A language model will write what a ripping alert means and summarise a case file, from the evidence Shiva already holds: the signals, the minutes, the counts. Every number in the text is checked against the record before it is shown, and the fixed sentences of today remain when it cannot be.
-
2
A baseline of each tenant's normal viewing.
Today's thresholds are the same for everyone. They will be learned from how each tenant's own viewers watch, so an unusual viewing stands out sooner and an ordinary one is left alone. An alert will still name the rule it tripped.
-
3
Reading the visible mark from a damaged copy.
When a copy is too degraded for the forensic mark, the visible text will be read from its frames and compared with the texts the tenant actually issued. Visible text can be faked to point at someone else, so this gives the investigator a lead, never a name.
-
4
Recognising a tenant's title in a file or a link.
A fingerprint of each title will let Shiva tell which video a suspect file is and where in it the clip starts, even re-encoded or cropped. That is what the detector needs to read the marks from a copy it cannot line up today.
It will run inside Shiva's own cloud boundary: a tenant's video and viewer references are not used to train a model and do not go to a public AI service.