For the last couple of decades, people have been highly motivated to get around network censorship, and willing to pay to do it. In China, VPN use is probably a few percent of internet users. That is still tens of millions of people, and a grey market on the order of $1 billion annually. Censors spend almost as much. Russia has budgeted about $100 million / year on it’s filtering system, and estimates for the Great Firewall’s network infrastructure run from $100 to $700 million a year.
That spending reflects a particular bargain. Historically, censorship regimes have been built around a simple cost-benefit analysis:
- Friction shifts behavior. Making information harder to reach changes what a population reads and believes, even if it doesn’t stop everyone.
- Collateral damage is expensive. Blocking too broadly hurts the commerce and economy the regime depends on.
The result is a deliberately leaky system. Motivated individuals can pay for a VPN or configure their way around restrictions, and the censor has accepted that. A more targeted system would cost more, cause more collateral damage, and do little to further the regime’s interests overall.
That bargain is starting to break. As access policies become cheap to update automatically and to apply at a finer grain, censorship will be constrained less by the costs we’ve traditionally assumed.
The cost curve bends
Targeting is getting cheaper. Identifying unwanted traffic, and even unwanted-but-obfuscated traffic, is getting easier. In 2021 the Great Firewall began passively detecting and blocking fully encrypted traffic in real time, something previously thought to require active probing. This changes what a censor can aim for. The goal no longer has to be a statistical shifts in behavior: actually cutting off access to samizdat is within reach.
The more important change is that costs no longer have to be spread evenly. If the average per-user cost of enforcement is low, the same budget can concentrate far more cost on a small slice of the population: the journalists, organizers, and technically capable users who matter most to the censor. Most people will never notice the difference. The few who are targeted will face a much harder adversary than the one we design for today.
This flips how we think about circumvention tools. We tend to measure success by aggregate numbers, such as how many users connect and how often the median connection works. In a world of precision censorship, those numbers say little. What matters is whether a tool still works for the targeted user.
Policy for one
Services are already applying LLM-driven decision flows to manage network traffic policy. This isn’t yet cost-effective at the backbone, but the techniques are getting there quickly. Flows that don’t match an existing policy can be held for analysis and decided on with the surrounding context. The real cost is gathering enough related context about a source IP cheaply, and that is exactly the cost that triage over time drives down.
For a targeted user, then, we should assume that all relevant context is stored and available to the censor when it decides what to do with their traffic. That assumption undercuts most of today’s transports:
| Transport | What gives it away |
|---|---|
| VPN | Enumerable IPs and ASes that can be blocked by policy |
| Look-like-nothing (obfs4, Shadowsocks) | All traffic to one IP, in no recognizable protocol |
| Domain fronting | Heavy traffic with no matching application fingerprint |
| WebRTC (Snowflake) | The surrounding metadata: which STUN and signaling traffic accompanies the flow |
Each of these holds up reasonably well against a censor looking at flows in aggregate. None holds up well against one looking at a single user’s full history.
That gives us a sharper design requirement. It isn’t enough for traffic to be unclassifiable in isolation. Traffic has to remain plausible given everything the censor knows about the user.
Confused deputies
There are two broad ways to meet that requirement. A user can make their traffic genuinely legitimate, or they can make sure the censor never sees enough of it to judge.
Becoming a legitimate user. A foreign business can legitimately set up a data center to serve its customers. If that business is actually run by the censored user, jurisdictional arbitrage comes into play. The traffic is real business traffic, consistent with a real history, and blocking it carries real costs.
Relaying for each other. Users can exchange and relay traffic, or encrypted fragments of it. When the traffic leaving any one node isn’t the full picture, there is no complete history to judge, and reconstructing one becomes much more expensive for the censor. Keeping up would also take further physical build-out of its network.
Both are structurally expensive ways to keep the censor a confused deputy: from its position in the network, it can’t see enough to make a correct policy decision.
Towards resilience
Neither strategy works well with today’s infrastructure. Each needs building blocks we don’t yet have.
Legitimacy needs circuit reservations. Legitimate-looking infrastructure is only useful if it can be acquired and replaced faster than the censor can profile it. This summer Cloudflare announced a Monetization Gateway that lets any resource behind it be paid for per request over x402, by any caller, with no account or prior relationship. A generalized protocol along those lines would let local networking find and pay for working routes on its own. Today we set up VPSes and VPNs on human timescales, with human contracts. Network conditions already change faster than most infrastructure exposes any way to adapt, and staying ahead of a targeting censor depends on closing that gap.
Distribution needs coordination without trust. Relaying only works if users can find each other and cooperate without exposing themselves. In encrypted messaging, deniability means a transcript could have been produced by either party alone. That makes it useless as proof of who said what, unlike leaked emails, where DKIM signatures have been used to establish provenance. Work on multi-party OTR showed that the property doesn’t carry over cleanly to groups. Once several participants hold the same transcript, their agreement undermines deniability, whatever the protocol itself provides.
We need a related property for networks, and we don’t yet have a construction for it. How do mutually distrusting nodes on a local network self-assemble? How do they reach the point of relaying traffic for each other, or carrying out-of-band signaling, in a way that is deniable? The goal is that network traces can’t pin individual users as running subversive software.
Distribution also needs multipath that works. Splitting traffic across peers is only practical if the result is usable. We have protocols that support forms of mobility and multipath, but none we can yet point to as reliable, efficient, and fast. We need to iterate toward a protocol that spreads traffic across multiple participating machines, adapts as conditions change, and efficiently reassembles the original intent on the other side. The undifferentiated VPN-as-a-stream abstraction is the hardest case for this, and the abstractions that would make it easier are underdeveloped. Building at the QUIC level, where different 5-tuples can be owned by different nodes and connection migration is already supported, may allow much better iterations of this kind of system. Multipath QUIC is the natural starting point.
Leave a Reply