NaïveProxy uses Chrome'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 in NaïveProxy:

• Website fingerprinting / traffic classification: mitigated by traffic multiplexing in HTTP/2.
• TLS parameter fingerprinting: defeated by reusing Chrome's network stack.
• Active probing: defeated by application fronting, i.e. hiding proxy servers behind a commonly used frontend with application-layer routing.
• Length-based traffic analysis: mitigated by length padding.

The code consists of a patch series morphed and rebased onto every new Chrome release.

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

NaïveProxy uses Chrome's network stack. The traffic behavior intercepted by the censor is identical to regular HTTP/2 traffic between Chrome and standard Frontend (e.g. Caddy, HAProxy).

Frontend also reroutes unauthenticated users and active probes to a backend HTTP server, making it impossible to detect the existence of a proxy, like this: Probe ⟶ Frontend ⟶ index.html

Starting from v84, users can run a naïve fork of Caddy forwardproxy without the Naïve server.

See latest release. Linux, Windows, Mac OS, and various OpenWrt targets are supported.

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

Note: On Linux libnss3 must be installed before using the prebuilt binary.

On the server, build and run Caddy v2 with naïve fork of forwardproxy: (Golang v1.14+ required)

go get -u github.com/caddyserver/xcaddy/cmd/xcaddy
~/go/bin/xcaddy build --with github.com/caddyserver/forwardproxy@caddy2=github.com/klzgrad/forwardproxy@naive
sudo setcap cap_net_bind_service=+ep ./caddy

Then ./caddy start with the following Caddyfile (replace the example values accordingly):

:443, example.com
tls [email protected]
route {
  forward_proxy {
    basic_auth user pass
    hide_ip
    hide_via
    probe_resistance
  }
  file_server { root /var/www/html }
}

This is Caddy 2 syntax. It is not the same with Caddy 1.

Locally 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:[email protected]"
}

See USAGE.txt for more parameters in config.json. See also Performance Tuning.

If you don't like to download binaries, you can build NaïveProxy.

Prerequisites:

• Ubuntu (apt-get install): git, python2, ninja-build (>= 1.7), pkg-config, libnss3-dev, curl, unzip, ccache (optional)
• MacOS (brew install): git, ninja, ccache (optional)
• Windows (choco install): git, python2, ninja, visualstudio2017community. See Chromium's page for detail on Visual Studio setup requirements.

Build (output to ./out/Release/naive):

git clone --depth 1 https://github.com/klzgrad/naiveproxy.git
cd naiveproxy/src
./get-clang.sh
./build.sh

The scripts download tools from Google servers with curl. You may need to set a proxy environment variable for curl, e.g. export ALL_PROXY=socks5h://127.0.0.1:1080.

Their TLS stacks have distinct features that can be easily detected. TLS parameters are generally very informative and distinguishable. Most client-originated traffic comes from browsers, putting the custom network stacks in the minority.

Previously, Tor tried to mimic Firefox's TLS signature and still got identified and blocked by firewalls, because that signature was of an outdated version of Firefox and the firewall determined the rate of collateral damage would be acceptable. If we use the signature of the most commonly used browser the collateral damage of blocking it would be unacceptable.

Any languages can be used for high performance architectures, but not all architectures have high performance.

Go, Node, etc. make it easy to implement a 1:1 connection proxy model, i.e. creating one upstream connection for every user connection. Then under this model the goal of performance is to reduce overhead in setting up each upstream connection. Toward that goal people start to reinvent their own 0-RTT cryptographic protocols (badly) as TLS goes out of the window because it either spends take several round trips in handshakes or makes it a pain to set up 0-RTT properly. Then people also start to look at low level optimization such as TCP Fast Open.

Meanwhile, Google has removed the code for TCP Fast Open in Chromium all together (they authored the RFC of TCP Fast Open in 2014). The literal reason given for this reversal was

We never enabled it by default, and have no plans to, so we should just remove it. QUIC also makes it less useful, and TLS 1.2 0-RTT session restore means it potentially mutates state.

And the real reason Google never enabled TCP Fast Open by default is that it was dragged down by middleboxes and never really worked. In Linux kernel there is a sysctl called tcp_fastopen_blackhole_timeout_sec, and whenever a SYN packet is dropped, TCP Fast Open is blackholed for this much time, starting at one hour and increasing exponentially, rendering it practically useless. Today TCP Fast Open accounts for 0.1% of the Internet traffic, so using it actually makes you highly detectable!

It was obvious to Google then and is obvious to us now that the road to zero latency at the cost of compromising security and interoperability is a dead end under the 1:1 connection model, which is why Google pursued connection persistence and 1:N connection multiplexing in HTTP/2 and more radical overhaul of HTTP/TLS/TCP in QUIC. In a 1:N connection model, the cost of setting up the first connection is amortized, and the following connections cost nothing to set up without any security or stability compromises, and the race to zero connection latency becomes irrelevant.

Complex, battle-tested logic for connection management was implemented in Chromium. The same thing is not so easy to do again from scratch with the aforementioned languages.

Because the first rule of cryptography is: Don't roll your own cryptos.

If you do roll your own cryptos, see what happened with Shadowsocks. (Spoiler: it encrypts, but doesn't authenticate, leading to active probing exploits, and more exploits after duct-tape fixes.)

You may have wondered why not use Chrome directly if NaïveProxy reuses Chrome's network stack. The answer is yes, you can. You will get 80% of what NaïveProxy does (TLS, connection multiplexing, application fronting) without NaïveProxy, which is also what makes NaïveProxy indistinguishable from normal traffic. Simply point your browser to Caddy as an HTTP/2 or HTTP/3 forward proxy directly.

But this setup is prone to basic traffic analysis due to lack of obfuscation and predictable packet sizes in TLS handshakes. The bane of "TLS-in-TLS" tunnels is that this combination is just so different from any normal protocols (nobody does 3-way handshakes twice in a row) and the record sizes of TLS handshakes are so predictable that no machine learning is needed to detect it.

The browser will introduce an extra 1RTT delay during proxied connection setup because of interpretation of HTTP RFCs. The browser will wait for a 200 response after a CONNECT request, incuring 1RTT which is not necessary. NaïveProxy does HTTP Fast CONNECT similar to TCP Fast Open, i.e. send subsequent data immediately after CONNECT. Also, you may have to type in the password for the proxy everytime you open the browser. NaïveProxy sends the password automatically.

Thus, traffic obfuscation, HTTP Fast CONNECT, and auto-authentication are the crucial last 20% provided by NaïveProxy. These can't be really achieved inside Chrome as extensions/apps because they don't have access to sockets. NaïveProxy extracts Chromium's network stack without all the other baggage to build a small binary (4% of a full Chrome build).

But if you don't need the best performance, and unobfuscated TLS-in-TLS somehow still works for you, you can just keep using Caddy proxy with your browser.

Take Cloudflare for example. https://www.cloudflare.com/terms/ says: "Use of the Service for serving video (unless purchased separately as a Paid Service) or a disproportionate percentage of pictures, audio files, or other non-HTML content, is prohibited." Proxying traffic is definitely prohibited by the terms in this context.