Home

Awesome

NaïveProxy build workflow

NaïveProxy uses Chromium's network stack to camouflage traffic with strong censorship resistence and low detectablility. Reusing Chrome's stack also ensures best practices in performance and security.

The following traffic attacks are mitigated by using Chromium's network stack:

Architecture

[Browser → Naïve client] ⟶ Censor ⟶ [Frontend → Naïve server] ⟶ Internet

NaïveProxy uses Chromium's network stack to parrot traffic between regular Chrome browsers and standard frontend servers.

The frontend server can be any well-known reverse proxy that is able to route HTTP/2 traffic based on HTTP authorization headers, preventing active probing of proxy existence. Known ones include Caddy with its forwardproxy plugin and HAProxy.

The Naïve server here works as a forward proxy and a packet length padding layer. Caddy forwardproxy is also a forward proxy but it lacks a padding layer. A fork adds the NaïveProxy padding layer to forwardproxy, combining both in one.

Download NaïveProxy

Download here. Supported platforms include: Windows, Android (with NekoBox), Linux, Mac OS, and OpenWrt (support status).

Users should always use the latest version to keep signatures identical to Chrome.

Build from source: Please see .github/workflows/build.yml.

Server setup

The following describes the naïve fork of Caddy forwardproxy setup.

Download here or build from source:

go install github.com/caddyserver/xcaddy/cmd/xcaddy@latest
~/go/bin/xcaddy build --with github.com/caddyserver/forwardproxy=github.com/klzgrad/forwardproxy@naive

Example Caddyfile (replace user and pass accordingly):

{
  order forward_proxy before file_server
}
:443, example.com {
  tls me@example.com
  forward_proxy {
    basic_auth user pass
    hide_ip
    hide_via
    probe_resistance
  }
  file_server {
    root /var/www/html
  }
}

:443 must appear first for this Caddyfile to work. See Caddyfile docs for customizing TLS certificates. For more advanced usage consider using JSON for Caddy 2's config.

Run with the Caddyfile:

sudo setcap cap_net_bind_service=+ep ./caddy
./caddy start

See also Systemd unit example and HAProxy setup.

Client setup

Run ./naive with the following config.json to get a SOCKS5 proxy at local port 1080.

{
  "listen": "socks://127.0.0.1:1080",
  "proxy": "https://user:pass@example.com"
}

Or quic://user:pass@example.com, if it works better. See also parameter usage and performance tuning.

Third-party integration

Notes for downstream

Do not use the master branch to track updates, as it rebases from a new root commit for every new Chrome release. Use stable releases and the associated tags to track new versions, where short release notes are also provided.

Padding protocol, an informal specification

The design of this padding protocol opts for low overhead and easier implementation, in the belief that proliferation of expendable, improvised circumvention protocol designs is a better logistical impediment to censorship research than sophisicated designs.

Proxy payload padding

NaïveProxy proxies bidirectional streams through HTTP/2 (or HTTP/3) CONNECT tunnels. The bidirectional streams operate in a sequence of reads and writes of data. The first kFirstPaddings (8) reads and writes in a bidirectional stream after the stream is established are padded in this format:

struct PaddedData {
  uint8_t original_data_size_high;  // original_data_size / 256
  uint8_t original_data_size_low;  // original_data_size % 256
  uint8_t padding_size;
  uint8_t original_data[original_data_size];
  uint8_t zeros[padding_size];
};

padding_size is a random integer uniformally distributed in [0, kMaxPaddingSize] (kMaxPaddingSize: 255). original_data_size cannot be greater than 65535, or it has to be split into several reads or writes.

kFirstPaddings is chosen to be 8 to flatten the packet length distribution spikes formed from common initial handshakes:

Further reads and writes after kFirstPaddings are unpadded to avoid performance overhead. Also later packet lengths are usually considered less informative.

H2 RST_STREAM frame padding

In experiments, NaïveProxy tends to send too many RST_STREAM frames per session, an uncommon behavior from regular browsers. To solve this, an END_STREAM DATA frame padded with total length distributed in [48, 72] is prepended to the RST_STREAM frame so it looks like a HEADERS frame. The server often replies to this with a WINDOW_UPDATE because padding is accounted in flow control. Whether this results in a new uncommon behavior is still unclear.

H2 HEADERS frame padding

The CONNECT request and response frames are too short and too uncommon. To make its length similar to realistic HEADERS frames, a padding header is filled with a sequence of symbols that are not Huffman coded and are pseudo-random enough to avoid being indexed. The length of the padding sequence is randomly distributed in [16, 32] (request) or [30, 62] (response).

Opt-in of padding protocol

NaïveProxy clients should interoperate with any regular HTTP/2 proxies unaware of this padding protocol. NaïveProxy servers (i.e. any proxy server capable of the this padding protocol) should interoperate with any regular HTTP/2 clients (e.g. regular browsers) unaware of this padding protocol.

NaïveProxy servers and clients determines whether the counterpart is capable of this padding protocol by the presence of the padding header in the CONNECT request and response respectively. The padding procotol is enabled only if the padding header exists.

The first CONNECT request to a server cannot use "Fast Open" to send payload before response, because the server's padding capability has not been determined from the first response and it's unknown whether to send padded or unpadded payload for Fast Open.

Changes from Chromium upstream

Known weaknesses