InvisiCloud

Technical details · Built on TLShare

The key never meets the ciphertext.

InvisiCloud splits the server end of a standard TLS connection between two independently operated parties. Neither ever sees your data in plaintext, and your client doesn’t change.

ciphertext Payload Gateway Has the records, never the key
keystream Key Server Has the key, never the records
plaintext Inside SMPC only Never reassembled in one place

The core idea

Every TLS record is already split in two.

AES-GCM encrypts by XORing a keystream onto the data. The ciphertext on the wire and the keystream from the session key are two halves of every record. Each half alone looks random. Together they are the plaintext.

In cryptographic terms they are XOR secret shares, exactly what secure multi-party computation (SMPC) takes as input. TLShare only has to keep the halves apart, and it never joins them outside SMPC.

Five bytes of a TLS record: ciphertext XOR keystream equals the plaintext “Hello”.
Ciphertext Payload Gateway 9f 3a c1 07 41
Keystream Key Server d7 5f ad 6b 2e
Plaintext inside SMPC 48 H 65 e 6c l 6c l 6f o

Other approaches split the key. TLShare keeps the key whole and keeps the ciphertext away from it.

Handshake
Plain TLS
The Key Server completes it alone. No SMPC.
Into shares
1 XOR / byte
Plus one AES-CTR keystream. Nothing else.
Client
Unmodified
Any standard TLS client. No agent, no SDK.

The protocol

One connection, three phases.

The Payload Gateway receives every byte but holds no key. After the handshake, the Key Server holds the key but sees only metadata.

Client
Payload Gateway
Key Server
01 Handshake · Plain TLS, no SMPC
TLS 1.2 / 1.3 handshake, relayed byte for byte
Holds the session key
02 Request · TLS → shares
Encrypted records: kept, not forwarded
Lengths, sequence numbers
Computes the AES-GCM keystream
ciphertext ⊕ keystream, computed inside SMPC
03 Response · Shares → TLS
SMPC outputs two shares of the response
Its share ⊕ keystream
⊕ own share = ciphertext
Standard TLS record
One connection in server-side mode. The Key Server never receives application ciphertext; the Payload Gateway never receives the key. Every byte sent to the client passes through this joint step.

TLShare also runs in reverse: in client-side mode the two parties jointly act as the TLS client toward an unmodified server, with the same mechanism.

Security model

Who sees what, and what we assume.

What the Payload Gateway and the Key Server can each see
Data Payload Gateway Key Server
TLS private key and session keys Never sees Sees
Encrypted application records Sees Never sees
Plaintext Only inside SMPC Never sees Never sees
Record lengths, sequence numbers, timing Accepted leakage Sees Sees
  1. A1

    No collusion

    The Payload Gateway and the Key Server are run by different entities: separate organizations, separate teams, or one team on two hyperscalers so that neither cloud has to be trusted. As long as one of them follows the protocol, nobody learns plaintext.

  2. A2

    Adversary model

    Semi-honest by default. A protocol extension adds active security: the AES-GCM tag is verified jointly and the shares become actively secure. Which one fits depends on the application.

  3. A3

    Detectable forgery

    In semi-honest mode the Payload Gateway learns the GHASH key to verify tags, which would let it alter records. It commits to each MAC before it receives that key, so tampering is detectable.

  4. A4

    No bypass

    The Key Server accepts no direct connections. Repointing the domain to itself would be publicly visible in DNS.

Measured cost

About one round trip per message.

Because the handshake needs no SMPC, the overhead is mostly network distance between the Payload Gateway and the Key Server.

per warm message
+1.1 ms
32 B to 16 KiB: 5.4 ms vs 4.3 ms with plain TLS
plain-TLS throughput
1.02×
4 connections: 290 vs 284 MiB/s, bounded by the link
per handshake
+0.24 ms
TLS 1.3 on loopback, before network round trips

Where the extra 1.1 ms goes (warm message, median)

One round trip between the Payload Gateway and the Key Server, plus about 0.4 ms of computation.

Plain TLS 4.3 ms TLShare ≈ 5.4 ms
  • Plain TLS 4.3 ms
  • Gateway ↔ Key Server 0.73 ms
  • TLShare work 0.4 ms
Show throughput table
Transfer Plain TLS TLShare Ratio
64 MiB, TLS 1.3, 1 connection 199 MiB/s 153 MiB/s 0.77
64 MiB, TLS 1.3, 4 connections 284 MiB/s 290 MiB/s 1.02
64 MiB, TLS 1.2, 1 connection 172 MiB/s 114 MiB/s 0.66
256 MiB, TLS 1.3, 1 connection 181 MiB/s 179 MiB/s 0.99
256 MiB, TLS 1.2, 1 connection 197 MiB/s 119 MiB/s 0.60

Preliminary medians. AWS Frankfurt, Payload Gateway and Key Server in different availability zones (0.73 ms round trip), client on Hetzner (4.4 ms). 2,056 messages each way, all decoded correctly. Single-connection throughput is 0.60–0.99× of plain TLS; with four connections the network link, not TLShare, is the limit.

Deployment & scope

What it needs, what it supports.

Deployment

Operators
Managed setup: utilacy runs the Payload Gateway, you or a trusted partner run the Key Server.
DNS
The Key Server owns the domain and keeps the private key. A/AAAA records point to the Payload Gateway.
Certificates
DNS-01 challenge, or HTTP-01 challenges forwarded to the Key Server.
Clients
Any standard TLS client, unchanged.

Protocol scope

TLS
TLS 1.2 and 1.3, AES-GCM cipher suites only.
Applications
Any protocol over TLS: HTTP, REST, FTPS, SMTP. InvisiCloud Objects speaks S3; the InvisiCloud Files web app talks to it over HTTPS.
Disabled
Session resumption and 0-RTT (enforced). Session tickets off by default.
Unsupported
TLS 1.3 KeyUpdate: the connection fails at the next record. Alerts pass through as shares.

Compare approaches

Different approaches. Different trade-offs.

Common security architectures prioritize confidentiality, compatibility, and control differently.

Advantage Trade-off Higher effort or limitation
ApproachConfidentialityTrust & controlCompatibilitySetup complexity
Server-side encryption Trade-off: Provider may access plaintext Higher effort or limitation: Trust concentrated in provider Advantage: Native integration Advantage: Low
Client-side encryption Advantage: Encrypted before upload Trade-off: Clients hold plaintext & keys Higher effort or limitation: Client changes required Higher effort or limitation: Per-application integration
Encryption gateways Trade-off: Gateway sees plaintext Trade-off: trusted gateway operator Advantage: Supported apps unchanged Trade-off: Gateway integration
Confidential computing Advantage: Protected inside enclave Trade-off: Hardware & attestation trust Trade-off: Workload adaptations Higher effort or limitation: Enclave integration
InvisiCloud Advantage: Neither party alone can access plaintext Advantage: Control split across independent operators Advantage: Existing applications & storage workflows unchanged Trade-off: Gateway deployment & routing

Ratings describe typical architectures; effort and protection depend on deployment. InvisiCloud reflects the target architecture and planned capabilities. See Technical Evaluation for current availability.

At the protocol level

TLShare compared with other ways of splitting TLS.

Approach What is split Handshake Result
Keyless SSL Only the private-key operation Standard The edge still gets session keys and plaintext
TLSNotary, DECO, zkTLS The client’s session key Two-party computation Client side only
TLShare Ciphertext from key Standard, no SMPC No single party sees plaintext; server or client side

Understand → Evaluate → Test

Ready to evaluate your workload?

Verify your required interfaces, setup, security requirements and performance targets.

Technical FAQ

Your technical questions, answered.

Have a question about your workload? Contact us