What the sources actually say

The official documentation defines the DPI desync attack as follows: take the original message, modify it and add fake information in such a way that the server OS accepts only the original data, while the DPI either cannot reconstruct the original message or sees something it cannot identify as a prohibited request.

The toolkit behind this is the nfqws --dpi-desync parameter. The readme lists the available modes: synack, fake, fakeknown, rst, rstack, hopbyhop, destopt, ipfrag1, multisplit, multidisorder, fakedsplit, hostfakesplit, fakeddisorder, ipfrag2, udplen and tamper. Modes can be combined, and additional parameters control split positions, fake payloads, TTL values and fooling methods.

How the main mode families work, step by step

  1. Fake-based modes (fake, fakeknown, rst, rstack) send generated packets carrying false information for the DPI. The readme stresses the key constraint: fakes must either never reach the server or be rejected by it, otherwise the TCP connection would break.
  2. Segmentation modes (multisplit, fakedsplit, hostfakesplit) divide the request at chosen positions so the DPI never sees the full prohibited pattern in one packet. multidisorder and fakeddisorder send the segments out of order.
  3. Fragmentation modes (ipfrag1, ipfrag2) do the same at the IP layer, and hopbyhop/destopt add IPv6 extension headers the DPI may fail to walk through.
  4. udplen changes UDP packet length, and tamper applies other protocol-level modifications documented in the readme.
  5. To make the server reject fakes while the DPI still reads them, the documentation lists fooling methods: bad checksums (badsum), bad sequence numbers (badseq), low TTL so fakes expire before the server, md5sig, timestamps (ts) and others, each with documented limitations.
The readme documents real limits for each fooling method: badsum packets are dropped by many home routers running Linux NAT, md5sig typically works only against Linux servers, and TTL-based fakes need per-ISP tuning. This is why no single mode is universally correct.

Why fake and modified traffic is used at all

The documentation's reasoning is that a passive DPI must decide from a single pass over the wire bytes whether a request is prohibited. If the byte stream the DPI assembles differs from what the server assembles, the DPI decides on the wrong data. Fakes, splits, disorders and sequence-number overlaps are all different ways to create that gap. The server-side operating system follows strict TCP/IP rules and recovers the original stream; the DPI takes shortcuts for performance, and those shortcuts are what the attack exploits.

The Flowseal community discussion #251 on YouTube troubleshooting shows what this looks like in practice: participants report user-specific DPI parameters, exchange strategy variants, and use blockcheck-style testing to find which combination of modes their particular ISP fails on. Those are community experiments tied to specific networks and versions, not universal recipes.

What is source-backed vs community-reported

  • Source-backed (official documentation): the desync definition, the full mode list, the constraint that fakes must not reach the server, the fooling methods and their documented limitations.
  • Source-backed (official documentation): split positions can be absolute offsets or relative markers such as method, host, endhost, sld, midsld and sniext.
  • Community-reported (Discussion #251): specific working parameter combinations for specific ISPs, and the practice of testing strategies instead of assuming one fits everyone. The linked discussion is a community thread and should be read as symptoms and hypotheses, not validated solutions.

When the reported workaround may not apply

A desync mode that one discussion participant reports as working can be useless or even harmful on another network: the documentation notes, for example, that datanoack can break NAT setups and that badseq fakes can also be ignored by the DPI itself. DPI equipment is updated over time, so any parameter set copied from an issue or discussion is historical evidence about one environment. Test on your own connection, and treat every posted strategy as a starting point rather than a fix.

Source notes

Every factual claim in this article is grounded in the sources below. Community reports are attributed and are not official recommendations. Last verified: 2026-09-30.