Skip to content

Latest commit

 

History

History
56 lines (39 loc) · 2.41 KB

File metadata and controls

56 lines (39 loc) · 2.41 KB

Threat Model

This is a short, honest summary. A full formal treatment will appear in the Diffuse white paper. Do not rely on Diffuse for high-risk use before reading it.

What Diffuse protects

Prompt content in transit. The client tokenizes and runs the first layers of the model locally. Only intermediate activations leave the machine, never the prompt tokens in clear text.

Inter-node confidentiality. Computation between nodes travels over an authenticated, end-to-end encrypted channel (X25519 key exchange, then ChaCha20-Poly1305 AEAD). There is no certificate authority; keys are bound to node identities.

Announcement integrity. Peer announcements in the gossip network are signed (Ed25519). Forged or tampered peer records are rejected.

Session unlinkability (partial). Each client request uses an ephemeral key-exchange keypair, so two requests are not trivially linkable by a shared key.

What Diffuse does NOT protect (by default)

Traffic metadata. The fact that you are talking, when, how much, and the shape of your traffic are observable. Diffuse does not run over Tor or a mixnet by default.

Network identity / IP. Your IP address is visible to the nodes you connect to. Masking it is left to the user (VPN, proxy, or their own Tor).

Intermediate activations on nodes. A node computes on the activations it receives. These are not the prompt in clear, but reconstructing content from intermediate activations is a known research area. The difficulty is tunable by how many layers the client keeps locally. It is made hard, not proven impossible.

Mid-pipeline plaintext. Computation on each node happens in the clear in memory (no trusted execution environment or homomorphic encryption).

Content self-revelation. If your prompt names you, no system can un-say it.

Sybil and reputation attacks. There is no anti-Sybil mechanism yet. A single actor can run many nodes.

Adversaries considered

Diffuse aims to protect against a curious node operator and a passive network observer with respect to prompt content. It does not currently defend against a global passive adversary (traffic analysis across the whole network) or a well-resourced actor performing activation inversion.

Summary

Diffuse raises the cost of surveilling what you think, within a defined perimeter. It is not anonymity, and it is not unconditional secrecy. Know the boundary before you cross it.