6 min read
Classic DPI works by rules: an engineer describes a protocol's signature and the equipment looks for a match. Machine-learning filtering works differently — a model is trained on traffic samples and classifies connections by a combination of statistical features for which nobody wrote an explicit rule. The practical difference: a rule can be evaded by changing one feature; a classifier looks at dozens at once.
The direction is documented in reports by human-rights and journalistic organisations: by February 2026 more than 450 VPN services had been blocked in Russia, and the 2025–2027 filtering programme contains a separate line item for AI-assisted traffic analysis (Human Rights Watch, March 2026; Zona Media, April 2026). The total budget of the equipment modernisation programme is around 60 billion roubles (CNews, March 2026).
What the verifiable sources do not contain is concrete effectiveness figures. VPN vendors' blogs carry claims like "97 % detection accuracy even for obfuscated protocols", but those are marketing materials with no methodology and no independent verification. We do not reproduce such numbers: a vendor's statement about how well its competitors get caught is not data.
The working formulation is this: the use of machine learning in filtering is confirmed as a direction and as a budget line; there are no publicly verifiable measurements of its effectiveness.
| | Rule (signature) | Classifier (ML) | |---|---|---| | What it describes | a specific feature: byte order, a handshake field | the distribution of dozens of features at once | | Who composes it | an engineer, by hand | training on labelled samples | | How it is evaded | change the described feature | you must change the whole statistical profile | | Errors | almost no false positives | false positives are unavoidable | | Update speed | edit the rule | retrain the model |
The last two rows matter most for understanding practice.
A model outputs not "yes/no" but a probability. The operator picks the threshold, and that is a choice between two kinds of damage: raise it and you let some tunnels through; lower it and you start cutting ordinary traffic.
Hence the effect users see: odd disconnections on services nobody blocked. A banking app, a corporate video call, a large file download — by statistical features all of these can resemble a tunnel. This is precisely why behaviour-based filtering is almost always tuned conservatively: for an operator the cost of mass false positives outweighs the benefit of a few caught connections.
Symptoms and a diagnostic order are covered separately: what TSPU is and why a VPN does not work.
Changing the port. "Non-standard port" is one feature among dozens, and on its own it stopped deciding anything long ago.
Simple obfuscation. A layer that shuffles bytes removes a specific signature but leaves the statistics: packet lengths, pauses, flow direction. That is enough for a classifier — and the very fact of "an encrypted stream with no recognisable protocol" becomes a feature in itself.
A rare protocol. Here the effect is the opposite of what people expect: the rarer a mechanism, the more it stands out against the mass. For the same reason, switching ECH on in heavily filtered networks sometimes works against the user — SNI and ECH.
Not differing, but coinciding. The only approach resilient to statistical classification is not to hide traffic but to make it indistinguishable from what constitutes the bulk of it. A handshake with a real, popular site produces not a "masked tunnel" but a connection with the same fields, the same certificate and the same behaviour as millions of ordinary sessions. Cutting that class wholesale means cutting the ordinary web.
The mechanics are covered separately: Reality and VLESS in plain language; a comparison of the surviving protocols: Hysteria2, TUIC and AmneziaWG.
A supply of addresses. A classifier works per connection, but restrictions are applied to addresses. When a specific address falls under a rule, the matter is settled by switching, and there only the number of servers matters.
The service runs on VLESS with Reality masking: the handshake is performed with a real third-party site, so the connection has neither a signature of its own nor a statistical profile different from an ordinary HTTPS session. There are over a hundred servers — when one address is restricted, switching takes a second. What the service sees and stores: what data BessyConnect keeps.
An honest boundary: the race between masking and classification does not end in a one-off victory for either side. We do not promise that any particular method will work forever — we promise that masking is updated on the infrastructure side rather than by your hands.
Can a neural network identify a VPN from encrypted traffic? A classifier works not with content but with connection statistics: packet lengths, pauses, flow direction. Encryption does not conceal those features, so recognition by them is possible in principle.
How many VPN services are blocked in Russia? According to human-rights organisations, more than 450 as of February 2026. The exact figure depends on counting methodology and differs between sources.
Is it true that detection accuracy has reached 97 %? Such figures appear in VPN vendors' blogs with no methodology and no independent verification. Reports by research and human-rights organisations contain no verifiable measurements of effectiveness.
Does changing the port or simple obfuscation help? Against signature rules, partly; against statistical classification, barely: obfuscation removes a specific feature but leaves the traffic profile.
So what actually works? Not masking as "something indeterminate", but coinciding with mass traffic: a handshake with a real site and the behaviour of an ordinary HTTPS session, plus a supply of addresses to switch to.
Instagram won't open or keeps lagging? BessyConnect VPN on VLESS+Reality restores access to Instagram so your feed, stories, and Reels load block-free. More
BessyConnect is a VPN for Telegram: restore chats and calls when Telegram isn't working. The VLESS+Reality protocol bypasses blocks and DPI on all devices. More
BessyConnect is a VPN for WhatsApp that restores voice and video calls when they're blocked. The VLESS + Reality protocol beats DPI. Get it for iOS and Android. More
