middlebox-interference   Middlebox / TCP normalization

The RKS Global report documents a two-tier Russian censorship architecture: TSPU network-layer blocking (documented by Xue et al. 2024) at the ISP level, now supplemented by mandated app-layer VPN detection in the 30 most popular Russian Android apps. This layered approach means a circumvention tool that successfully bypasses TSPU at the network layer can still be detected and reported by the app layer, closing the gap that network-only circumvention leaves open.

2026-rks-russian-apps-vpn-detection §2, §7 deployment

Iran's censorship infrastructure began dropping all fragmented TLS ClientHello packets around late July 2026, rendering the tlshello fragmentation bypass technique non-functional. The blocking took effect approximately July 27, 2026, based on user-reported timelines in the net4people/bbs community report.

2026-spiderxss-tlshello-fragmentation-not Issue body (Jul 31 2026) detection

The server-side variant of the blind VPN inference attack—where an in/on-path adversary exploits predictable NAT assignment and tunnel routing semantics to inject spoofed packets indistinguishable from legitimate encrypted traffic—has remained unacknowledged and unmitigated across all tested platforms since its concurrent disclosure in 2019. Unlike the client-side variant, which received partial fixes from Google (CVE-2019-9461, CVE-2024-49734) and Apple (iOS 17.2.1), no vendor has proposed a viable remediation or claimed ownership of the server-side attack surface.

2026-tolley-architectural §2.1, §3.1, §3.4 detection

Application packaging and software supply chains in Russia's Android ecosystem can expand the effective TLS trust boundary beyond OS-level controls, enabling MITM attacks on TLS connections for apps that bundle or inherit Russian-controlled certificate authorities. This bypasses the assumption that the OS trust store constrains which CAs can intercept encrypted traffic.

2026-tolley-beyond-os-trust Abstract detection

After approximately four days, Iran placed Cloudflare IP ranges behind a NAT that performed full TCP reassembly with strict checksum and sequence-number validation; any packet with an incorrect checksum or out-of-window sequence number caused immediate connection teardown, eliminating both TTL-spoofing and sequence-desynchronization bypasses.

2026-tuleo1-internet-analysis-shutdown Issue body deployment

Two TCP-layer SNI-spoofing techniques briefly circumvented Iran's DPI for approximately four days: (1) sending a fake ClientHello with a TTL too low to reach the server but high enough to pass the censor, and (2) sending a fake ClientHello at a future TCP sequence number that the server would ignore but the stateless DPI would accept. Both were defeated once Iran placed Cloudflare IPs behind a stateful NAT that validated checksums and sequence numbers. Iran's SNI whitelist permitted only specific domains, hcaptcha.com among them.

2026-tuleo1-internet-analysis-shutdown Issue body (SNI spoofing methods paragraph) defense

Russian TSPU devices directly block ECH by dropping ClientHello messages that contain both an ECH extension and the outer SNI hostname "cloudflare-ech.com" — the static outer SNI Cloudflare advertises in all its ECH configurations. Blocking affects both TLS and QUIC. ECH connections to servers with Cloudflare ECH support but outside Cloudflare's official IP ranges are NOT blocked. TCP segmentation alone or TLS record fragmentation alone did NOT bypass TSPU ECH blocking, but combining both techniques did circumvent it. TSPU has also added TCP reassembly capabilities that defeat previously effective fragmentation-only bypasses.

2025-niere-encrypted §5, Figure 5 detection

The GFI's HTTP and HTTPS filters are now stateful (requiring initial SYN packet with matching sequence numbers) and have been activated on all TCP ports—not only standard ports 80 and 443 as reported by prior studies. This is a significant departure from previous work that found stateless HTTP/HTTPS blocking limited to standard ports. The HTTP filter injects a 403 Forbidden blockpage (not RST packets as used by the GFW), while HTTPS injects a single RST+ACK packet. The GFI also exhibits TCP non-compliance (not requiring a full three-way handshake to trigger filtering), enabling outside-in measurement without in-country servers.

2025-tai-irblock §2.1, §3.2 detection

The Henan Firewall is stateless in two exploitable ways: (1) it requires the TCP header to be exactly 20 bytes—enabling any TCP option (e.g., TCP Timestamps, which Windows disables by default) to bypass it entirely; (2) it does not perform TCP reassembly, so splitting a TLS ClientHello across two TCP segments such that the SNI extension straddles the boundary bypasses the censor. Both bypasses require only client-side changes and have already been implemented in Xray, GoodbyeDPI, and Shadowrocket. TLS record fragmentation (splitting the ClientHello across multiple TLS records within one TCP segment) also defeats both the Henan Firewall and the GFW, since neither performs TLS reassembly.

2025-wu-regional-censorship §4.3 / §6 defense

Longitudinal GFWeb data spanning 20 months shows the GFW actively patched previously-published evasion findings during the measurement period: overblocking bugs reported in academic papers were fixed, and fragmented-packet reassembly failures that researchers used to bypass blocking were corrected. This demonstrates that the GFW operator monitors published research and iterates on the system in response to disclosed vulnerabilities.

2024-hoang-gfweb Abstract, §5.4, §6 detection

The SAT Móvil app (Mexico's official tax service, 1M+ downloads) consistently fetches its 'Chat' page over cleartext HTTP, exposing citizen ID numbers (CURP, RFC), passwords, and tax documents to any in-path attacker. None of the four major Latin American telco apps (MiTelcel, MiTigo, MiClaro, MiMovistar) implement HSTS on SMS-delivered external links, making all of them uniformly vulnerable to SSL strip downgrade attacks.

2024-kujath-analyzing §7.2, Table 3 evaluation

China's Great Firewall showed anomalous inconsistency: 13 test vectors produced mixed outcomes—TCP RST injection on some executions and a clean server response on others—with circumvention rates between 10% and 35% across 100 executions per vector. The authors attribute this to heterogeneous GFW infrastructure components applying different HTTP parsing logic, a departure from the GFW's usual consistency.

2024-m-ller-turning §5.2 detection

Of 4,488 total HTTP Request Smuggling test vectors, 2,015 (44.9%) were accepted by at least one web server. CL*/TE vectors had a 99.0% acceptance rate (1,103/1,114); TE*/CL had 76.0% (859/1,130); CL/TE* had only 4.7% (53/1,130); and TE/CL* had 0%. Nginx 1.25.2 accepted 1,315 vectors while Apache 2.4.57 accepted only 11, reflecting HRS countermeasures added in Apache 2.4.25 and 2.4.52.

2024-m-ller-turning §5.1 / Table 1–4 evaluation

Web security vulnerabilities whose exploitation depends on parser divergence between two co-located systems are structurally isomorphic to censorship circumvention attacks, where the censor acts as the frontend parser and the destination server as the backend. The authors demonstrated this by directly converting all HRS test vectors from prior security research into circumvention probes with no modification, showing that censorship-circumvention techniques can be systematically constructed from existing vulnerability corpora.

2024-m-ller-turning §3, §8 (Conclusions) defense

The root cause of port-shadow vulnerabilities is that connection-tracking frameworks maintain five shared, globally-accessible resources across all VPN clients on the same server. The paper's formal model identifies these as: the conntrack table, the NAT table, the port space, the routing table, and the ARP/neighbor cache. Any of these shared resources can be used as a side-channel. Bounded model checking confirmed that enforcing strict process isolation around all five resources eliminates the attack surface.

2024-mixon-baca-snitch §4 (Formal Model), §6 (Mitigations) detection

The "port shadow" exploit abuses five shared, limited resources in Linux conntrack/Netfilter (and analogous frameworks in BSD, Windows) to let an off-path attacker intercept or redirect encrypted VPN traffic, de-anonymize a VPN peer's source IP, or portscan a peer hidden behind a VPN server — all without compromising the VPN's cryptographic layer. Four concrete attacks are demonstrated; formal model checking with bounded model checking verified six process-isolation mitigations that prevent the shared-resource collision.

2024-mixon-baca-snitch Abstract, §3, §5 detection

Stateful firewalls used as censorship middleboxes exhibit counter-intuitive implementation behaviors: FW-3 forwards ACK packets before a TCP handshake is initiated, and FW-1 actively spoofs RST packets in response to unsolicited traffic to thwart evasion attempts. These vendor-specific quirks create or close evasion opportunities that are invisible to rule-verification tools and not predictable from policy documentation alone.

2024-moon-pryde §1 Introduction, Findings 3–4 detection

Evasion attacks generated against one firewall-deployment combination do not transfer well to other settings: a deployment-agnostic approach (used by censorship circumvention tools) fails to generate effective attacks across diverse victim stacks and attacker capabilities. Pryde's deployment-aware, modular workflow finds successful attacks across configurations with and without insider threats, and against multiple attacker success criteria (data delivery vs. victim ACK vs. attacker receipt of ACK).

2024-moon-pryde §1 Introduction (Finding 7), §3.2 evaluation

TCP-compliant packet alphabets are insufficient for modeling stateful firewall evasion. Including non-TCP-compliant traffic — specifically flipped-direction SYNs, out-of-window seq/ack numbers, and packets that form a parallel TCP connection in the reverse direction — is what unlocks discovery of deep attack paths. Prior model-inference work (Alembic) that restricted itself to compliant sequences produced models incapable of generating any of the 6,000+ attacks Pryde found.

2024-moon-pryde §4.1–§4.2 detection

Pryde generates more than 6,000 successful and unique evasion attacks against 4 popular stateful firewalls, which is 2–3 orders of magnitude higher than censorship circumvention algorithms (e.g., Geneva) and black-box fuzzing. The gap arises because circumvention tools only uncover shallow evasion sequences and cannot systematically explore the full attack-state space.

2024-moon-pryde §1 Introduction, Finding 1 evaluation

In Brunei, censorship is confined to AS10094, which serves approximately 70% of the country's Internet users. The censor injects RST packets bearing a distinctive fingerprint — the censored query's IP ID field — in response to HTTP requests containing censored Host headers, and censors on all ports without residual censorship. A SYN followed immediately by a PSH+ACK with a censored payload is sufficient to trigger blocking without a completed TCP handshake.

2023-nourin-detecting §3 detection

Censoring middleboxes' TCP non-compliance — specifically, their willingness to censor bidirectionally without completing the three-way handshake — enables external vantage points outside a censoring country to trigger and measure censorship without any local endpoint participation. The approach requires only a confirmed censored domain per AS, evidence of bidirectional censorship, and minimal residual censorship.

2023-nourin-detecting §2 evaluation

Geneva — originally designed to evolve censorship-evasion packet sequences — was repurposed by inverting its fitness function to discover censorship-triggering packet sequences instead. Training against non-responsive IP addresses allows Geneva to attribute all responses to middleboxes, enabling fully automated discovery of triggering strategies without any endpoint cooperation.

2023-nourin-detecting §2 evaluation

Tajikistan routes virtually all national egress and ingress traffic through a single state-run AS (AS51346, Tojiktelecom) under a 2016 national decree, creating a centralized chokepoint. The censor injects RST+ACK packets with a unique 22-byte all-zero payload, censors on all ports, and requires two PSH+ACK packets containing the censored content before injecting — possibly modeling typical multi-resource HTTP browsing behavior.

2023-nourin-detecting §3 detection

Server-side censorship evasion strategies require zero client-side changes: clients bypass censorship without installing software or even being aware of the evasion, and this approach has been adopted in production tools including Psiphon's packetman. The packet manipulations exploit weaknesses in how censors track or tear down TCP connections, occurring entirely at the server during the three-way handshake.

2023-tran-crowdsourcing §1 Introduction defense

All existing automated server-side strategy discovery tools — Geneva, Alembic, and SymTCP — require researcher control of a client during training, even when the discovered strategies are deployed exclusively server-side. This dependency makes it infeasible to train against censors in networks where researchers cannot place a controlled machine.

2023-tran-crowdsourcing §1 Introduction evaluation

India's Airtel HTTP censor fails to reassemble TCP segments: padding any HTTP request to at least 1,449 bytes causes the IP+TCP overhead (52 bytes) to push the total past the Ethernet MTU of 1,500 bytes, forcing segmentation that the censor cannot handle and achieving 100% evasion. Kazakhstan requires the segmentation boundary to fall precisely between the Host header name and value (with two trailing spaces), rather than anywhere in the request.

2022-harrity-get §5.2 evaluation

Dominant failure modes differ systematically by country: China (AS45090) shows connect timeouts in 75% of DoT failures (IP-level blocking); Kazakhstan (AS48716) shows post-TLS-handshake timeouts in 72% of DoT failures (likely ACK or segment discard after handshake); Iran (AS197207) shows TLS handshake timeouts in 80% of DoT failures. Packet capture analysis confirmed that timeouts during and after the TLS handshake correspond to unacknowledged TCP segments, not connection resets.

2021-basso-measuring §V-F, §V-G, Table VIII evaluation

Internet-wide IPv4 scanning found 386,187 IP addresses yielding amplification factors ≥ 100× via TCP middlebox reflection, with 82.9% of responses from the top 1 million IPs confirmed as originating from on-path middleboxes rather than endpoints. Nation-state censorship infrastructure dominates: China's GFW alone accounts for approximately 154 million responding IP addresses sharing a 3× RST+ACK (54 bytes each) fingerprint.

2021-bock-weaponizing §5.3, §5.5, Table 4 evaluation

Censoring middleboxes respond to non-compliant TCP sequences because they must handle asymmetric routing and cannot rely on observing both sides of a connection. The hSYN; PSH+ACKi sequence elicited responses from 69.6% of 184 tested censoring middleboxes with a maximum amplification of 7,455×, while a lone PSH+ACK with no prior handshake elicited responses from 33.2% of middleboxes.

2021-bock-weaponizing §2, §3.3 detection

Nation-state censors produce characteristic TCP response fingerprints: China's GFW sends 3× RST+ACK (54 bytes each) from ~170 million IPs; Iran's infrastructure sends 402–405-byte FIN+PSH+ACK plus 54-byte RST+PSH+ACK from 8.6 million IPs (75.7% of responsive Iranian addresses); Saudi Arabia sends a 97-byte PSH+ACK plus 2× 1,354-byte PSH+ACKs at 18.9× amplification from 400,000+ IPs. Most nation-state censors produce less than 4× amplification due to compact block pages.

2021-bock-weaponizing §5.5, Table 4 detection

Routing loops within censoring infrastructure create effectively infinite TCP amplification: 53,041 of the top 1 million responding IP addresses showed routing loop behavior spanning 2,763 /24 prefixes. Two Russian ISP censorship systems with infinite routing loops continuously sent amplified traffic for approximately 6 days after a single 2-packet trigger sequence, and 6 GFW IP addresses in China sent data indefinitely.

2021-bock-weaponizing §5.6, §6 evaluation

Residual censorship—where a censor continues blocking all traffic on a 3- or 4-tuple after an initial censorship event—is active in China (HTTP: 90s 3-tuple RST injection; ESNI: 120–180s 3+4-tuple null routing), Iran (HTTP+SNI: 180s 4-tuple null routing, occasionally up to 5 minutes; protocol filter: 60s), and Kazakhstan (HTTP+SNI: 120s 4-tuple null routing). A December 2020 Quack scan found 3-tuple stateful disruption in 33 countries and null-routing censorship in 18, suggesting much broader applicability.

2021-bock-your §IV, Table I evaluation

All tested censors (China, Iran, Kazakhstan) can be triggered statelessly—without completing a TCP 3-way handshake—using a SYN with decremented sequence number followed by a PSH+ACK containing the forbidden payload. This stateless triggering enables fully off-path, source-spoofed attacks: an adversary with packet-spoofing capability can residually censor a victim pair they have no on-path access to.

2021-bock-your §IV, §V.A detection

Iran and Kazakhstan reset the residual censorship timer whenever the censor observes any matching packet from the victim, so TCP retransmissions from the victim's own stack inadvertently extend the blocking window far beyond the nominal 120–180s. China's HTTP residual censorship has only ~50% per-request reliability from some vantage points due to heterogeneous GFW middlebox load-balancing, but reliability plateaus near 100% after 7 repeated censorship triggers sent ahead of time.

2021-bock-your §IV (timer reset), §IV (reliability), Fig. 2 evaluation

The authors developed 'Aladdin,' a 10-step OONI-based measurement experiment that isolates SNI-based blocking (step 1), Host-header blocking (step 2), DNS injection (step 3), system-resolver vs. DoH discrepancy (steps 4–5), TLS interception (steps 6–8), and TLSv1.3-specific SNI dependency (step 10); this methodology exposed Vodafone's Allot TLS interception that OONI's Web Connectivity test had recorded only as a generic certificate error.

2021-ververis-understanding §3.8 evaluation

Vodafone (AS12357, AS12430, AS6739) deployed Allot-based TLS interception to block womenonweb.org: the system resolver returned a legitimate IP (67.213.76.19), but connecting to it triggered a forged certificate signed by Allot; disabling TLS certificate validation fetched the Vodafone blockpage, confirming a man-in-the-middle box rather than a redirect. OONI's standard Web Connectivity test recorded only a generic ssl_error:certificate verify failed and missed this entirely.

2021-ververis-understanding §3.7, §3.8.3 detection

A censor can identify Slitheen relay connections by observing that all packets in a suspected overt flow arrive in strict order while flows from the same source naturally exhibit out-of-order delivery: the relay station's traffic-server component reorders TCP segments to enable TLS record decryption, creating a statistically anomalous per-connection ordering pattern. The reordering buffer also increases per-packet round-trip times, providing a secondary timing signal.

2020-birtel-slitheen §4, §6 detection

China's GFW uses distinct, co-located censorship boxes—each with its own independent network stack implementation and bugs—for each application-layer protocol it censors. TCP-level strategies that exploit transport-layer bugs show dramatically different success rates per protocol: Strategy 1 (Simultaneous Open + Injected RST) achieves 89% for DNS but only 14% for HTTPS; Strategy 8 (TCP Window Reduction) achieves 100% for SMTP but only 2–3% for DNS, HTTP, and HTTPS. TTL-limited probes confirm all protocol boxes are co-located at the same network hop.

2020-bock-come §6, Table 2 detection

The paper identifies three distinct GFW resynchronization-state triggers with protocol-specific behavior: (1) a server payload on any non-SYN+ACK packet causes resync on the next SYN+ACK or client ACK-flagged packet for all protocols; (2) a server RST causes resync on the next client packet for all protocols except HTTPS; (3) a SYN+ACK with a corrupted acknowledgment number triggers resync only for FTP. Strategy 1's 50% per-attempt success rate for HTTP is confirmed to result from the 50% probability of the GFW entering the resynchronization state on an injected RST, consistent with Wang et al. [36].

2020-bock-come §5.1 detection

The paper presents 11 purely server-side censorship evasion strategies requiring zero client-side software, successfully bypassing censorship in China, India, Iran, and Kazakhstan across DNS-over-TCP, FTP, HTTP, HTTPS, and SMTP. All strategies manipulate only TCP handshake packets (primarily the SYN+ACK) and were verified against 17 versions of 6 client operating systems (Windows XP–Server 2018, MacOS, iOS, Android, Ubuntu, CentOS) with unmodified clients.

2020-bock-come §1, §5, Table 2 defense

TCP Window Reduction (Strategy 8)—reducing the SYN+ACK TCP window to 10 bytes and stripping wscale options, forcing the client to segment its request—achieves 100% evasion success against HTTP in India and Kazakhstan, 100% against HTTP and HTTPS in Iran, and 100% against SMTP in China, because none of these censors can reassemble TCP segments. The strategy is compatible with all 17 tested client OS versions when implemented without SYN+ACK payloads, making it the most broadly deployable server-side strategy found.

2020-bock-come §5.1, §5.2, §7, Table 2 defense

Using Geneva's genetic algorithm trained against Iran's live protocol filter, four evasion strategies achieving 100% success were discovered in under two hours: (1) injecting a fingerprint-matching PSH/ACK with a corrupt checksum before the real data; (2) sending two FIN packets before the SYN; (3) sending nine non-data-carrying packets (any flags, any seq/ack) during the handshake to exhaust the filter's per-flow packet limit; (4) a server-side variant that sends nine corrupted SYN+ACKs, inducing nine client RSTs before the real ACK, enabling fully unmodified clients to benefit.

2020-bock-detecting §5.2–§5.3 defense

Existing segmentation strategies effective against Iran's standard HTTP DPI can be counterproductive when the protocol filter is also active: if the first segment is fewer than 8 bytes, it fails the HTTP fingerprint check and trips the filter. However, segmenting such that the first segment is a valid HTTP fingerprint (≥8 bytes, well-formed verb + space) while splitting the Host: header into the second segment defeats both the protocol filter and the standard DPI censor simultaneously.

2020-bock-detecting §5.1 defense

The GFW's dominant exploitable discrepancy is accepting data packets whose TCP sequence number is ≤ the initial sequence number (ISN), while Linux rejects such packets as out-of-window. This single 'SEQ ≤ ISN' strategy accounts for the majority of the 3,152 successful evasion-packet cases against the GFW out of 4,587 total successful evasions.

2020-wang-symtcp §VIII.C detection

Snort contains two novel TCP Timestamp discrepancies versus Linux: it omits RFC 7323-mandated timestamp validation on RST packets in SYN_RECV state, and its PAWS TSval acceptance window is 'off by two' — a TSval of 0 or 0xffffffff following a packet with TSval 0x80000000 is accepted by Linux but rejected by Snort, enabling insertion-based evasion by crafting packets that fall in the divergent range.

2020-wang-symtcp §VIII.E detection

Snort interprets the TCP urgent pointer as the offset to the last byte of urgent data and discards all payload bytes before that offset, while Linux consumes only 1 urgent byte and leaves the remaining payload intact. Injecting a packet with the URG flag and the urgent-pointer offset pointing to an insignificant padding byte allows the full sensitive payload to reach the server while Snort strips it — a novel evasion strategy not previously reported.

2020-wang-symtcp §VIII.E detection

SymTCP uses selective symbolic execution over Linux's TCP implementation (S2E + KLEE) to enumerate all packet sequences reaching 47 binary-level accept or drop points from LISTEN to ESTABLISHED, then conducts differential testing against a blackbox DPI to confirm discrepancies; the open-sourced system requires no DPI source access and covers 37 of 47 drop points within the operationally relevant handshake window.

2020-wang-symtcp §IV, §VIII.A–B defense

SymTCP generated 56,787 candidate insertion/evasion packets in approximately one hour using concolic execution over Linux's TCP stack. Evaluating a sampled set of 10,000 test cases against real DPI systems yielded 6,082 evasions against Zeek, 652 against Snort, and 4,587 against the Great Firewall of China — discovering 14 novel evasion strategies beyond those found by prior manual approaches.

2020-wang-symtcp §VIII.C evaluation

Evasion strategies are strongly censor-specific: TCB Teardown strategies that achieve 80–96% against the GFW fail completely (0%) against Kazakhstan's HTTPS MITM; India's Airtel is defeated uniquely by a 'Stutter Request' (duplicating the PSH/ACK and replacing IP length to 64) at 100% success, which scores only 3% against the GFW. Geneva converged on distinct species for each censor within 4–8 hours of live training.

2019-bock-geneva Table 1, §5.1–5.4 evaluation

The GFW does not verify TCP checksums or validate RST flag combinations: Strategy 5 using the entirely invalid flag set FRAPUN with TTL 10 achieved 96% success. Separately, increasing the TCP data offset (dataofs) field to 10 in an insertion duplicate causes the GFW to reinterpret the beginning of the HTTP payload as TCP header bytes, preventing keyword detection and achieving 98% success (Strategy 2) — while the destination server discards the malformed packet.

2019-bock-geneva §5.2 Species 1 and Species 2 detection

Geneva's Segmentation species — fragmenting HTTP requests at the TCP layer without IP fragmentation, segment overlapping, or insertion packets — achieved 94–98% success against the GFW, 100% against India's Airtel ISP, and 100% against Kazakhstan's HTTPS MITM, making it the only strategy class effective across all three tested censors. These strategies require neither raw sockets nor root privilege.

2019-bock-geneva §5.2 Species 3: Segmentation defense

MultiFlow mitigates TLS termination attacks—where an adversary drops a connection after one data exchange—by having the client exfiltrate TLS session resumption information (219 bytes: 208-byte psk identity plus ticket metadata) to the decoy router. The decoy router can then resume a session with a different decoy host, establishing a new covert channel even if the original connection is severed, and amortizing per-session setup cost across multiple connections.

2018-manfredi-multiflow §3.3 defense

The paper enumerates at least eight distinct non-censorship motivations for server-side geo-blocking — economic sanctions, third-party liability (SESTA), copyright, GDPR compliance, security/fraud concerns, hosting costs, revenue optimization, and misconfiguration — each of which can produce the same observable signals (403 blockpages, DNS failures, TCP resets) as government censorship. Naive measurement methods that treat all location-based unavailability as censorship will produce systematic false positives.

2018-tschantz-bestiary §2, Table 1 policy

Of the 55 filters that inspected the HTTP Host header, 26 keyed only on the first Host header in a multi-Host request, 27 keyed only on the last, and only 2 examined both. Placing a benign Host header in the position the filter reads and the blocked URL in the other position bypassed the filter, and this divergence in behavior tracks RFC 7230's requirement to reject multi-Host requests with a 400 error — which none of the tested filters implemented.

2017-jermyn-autosonda §4.1 Mechanism detection

Among the 44 non-DNS filters, 11 did not reassemble TCP segments and 7 did not reassemble IP fragments before inspection, meaning a censored URL split across segment or fragment boundaries evaded detection. Five filters applied fragment/segment reassembly timeouts of under 2 seconds despite maintaining HTTP request state for more than 8.5 seconds, creating a window where a deliberately fragmented flow with artificial delay avoids inspection entirely.

2017-jermyn-autosonda §4.1 Mechanism detection

TCP segment splitting and out-of-order delivery evades DPI classification in the testbed, T-Mobile, and Iran, but fails against the GFC—which performs extensive packet validation and correctly reassembles reordered streams—and AT&T, which uses a transparent HTTP proxy that normalizes all traffic before inspection. Payload splitting to one byte in the first packet is sufficient to defeat packet-count-limited classifiers.

2017-li-lib-cdot-erate §4.3, Table 3 evaluation

lib·erate's TTL-limited inert packet insertion—sending a decoy packet with TTL set to expire at the middlebox but carrying a misclassifying payload—successfully evades classification in a carrier-grade testbed DPI device, T-Mobile's Binge On, and the Great Firewall of China, but fails against Iran's censor and AT&T (Table 3). When bilateral server support is available, inserting a single dummy packet at flow start evades classification in all four deployments.

2017-li-lib-cdot-erate §4.3, Table 3 evaluation

Of 2,134 tested sites, 229 (10.7%) were invalid for inbound blocking detection due to ingress filtering or network-origin discrimination; 431 additional sites were invalid for outbound blocking detection, of which 75% were Cloudflare-hosted and 7% Fastly-hosted because anycast topology prevents RST packets from returning to the originating anycast node.

2017-pearce-augur §V-E evaluation

The GFW evolved to create a TCB not only on SYN packets but also on SYN/ACK packets, and enters a 're-synchronization state' upon seeing multiple SYN packets, multiple SYN/ACK packets, or a SYN/ACK with an incorrect acknowledgment number. Once in this state, it re-synchronizes its TCB using the next client-to-server data packet or server SYN/ACK, invalidating prior TCB-creation evasion strategies that assumed the GFW used only the first SYN sequence number.

2017-wang-your §4 detection

Packets carrying an unsolicited TCP MD5 option header (RFC 2385) are silently ignored by modern Linux servers (kernel ≥ 2.6) that have not negotiated MD5 authentication, yet are accepted and processed by the GFW as normal packets that update its TCB. Crucially, none of the observed middleboxes dropped packets with MD5 options, making the MD5 header the most universally applicable insertion packet type — usable with any TCP flag (SYN, RST, or data) and immune to middlebox filtering.

2017-wang-your §5.3, Table 3 defense

Client-side middleboxes at every tested vantage point interfere with IP-layer evasion tactics: Aliyun (6/11 nodes) discards all IP fragments, while the Tianjin China Unicom node drops packets with wrong TCP checksums or no TCP flag. IP-layer discrepancies that survive routers (e.g., IP total-length > actual length) are still dropped by some middleboxes, making IP-layer manipulations unreliable across Chinese ISPs. TCP-layer manipulations are significantly more consistent across paths.

2017-wang-your §3, Table 2 evaluation

Castle structurally avoids all three covert-channel pitfalls identified by Geddes et al.: architecture mismatch is avoided by supporting both client-server and P2P modes; channel mismatch is avoided because RTS games implement application-layer reliability over UDP (matching proxied TCP requirements, unlike VoIP), blocking selective-drop denial-of-service attacks; content mismatch is avoided because legitimate RTS traffic has high natural variance driven by map, strategy, and player count.

2016-hahn-games §6.3 defense

The survey identifies 'soft censorship' — including throttling, packet-loss injection, and quality-of-experience degradation — as detected by only 2 of 13 surveyed platforms (rTurtle and UBICA) as of 2015. The paper explicitly flags this as a measurement gap, noting that soft censorship symptoms are indistinguishable from ordinary network congestion without ground-truth probes placed outside the censor's network.

2015-aceto-internet §3 / Table 2 detection

Marionette is the first programmable obfuscation system to simultaneously satisfy all five threat-model dimensions evaluated in Figure 2: resistance to blacklist DPI, whitelist DPI, statistical-test DPI, protocol-enforcing proxy traversal, and multi-layer traffic control, while sustaining throughput above 1 Mbps (up to 6.7 Mbps). Every prior system (obfs4, ScrambleSuit, SkypeMorph, StegoTorus, FTE, JumpBox, etc.) fails at least one dimension, most commonly stateful proxy traversal or statistical-feature control.

2015-dyer-marionette §2, Figure 2 defense

Format-Transforming Encryption (FTE) fails under proxy-induced ciphertext modification — a single character change causes decryption failure — while Marionette's probabilistic context-free grammar (CFG) templates tolerate header rewriting, connection multiplexing, and content alteration by intermediate proxies. Validated across 10,000 streams through Squid 3.4.9, achieving 5.8 Mbps downstream and 0.41 Mbps upstream goodput.

2015-dyer-marionette §7.3 defense

Rostelecom (AS12389) performed network-layer redirection of blacklisted traffic rather than DPI-based filtering: 40 of 343 Russian probes returned SSL certificates attributed to Russian ISPs (State Institute of Information Technologies, Rostelecom, Electron Telecom Network). The interference affected all protocols and ports holistically across Rostelecom's downstream peers, consistent with BGP-level false advertisements or forwarding rules rather than application-layer classification.

2014-anderson-global §4.2 detection

Within a single country mandate, different ISPs implement censorship with different filtering tools and mechanisms: Thailand's AS 9737 and AS 17552 use structurally distinct block-page templates (vector 17 is ~1,000 bytes using div layout; vector 8 is ~6,000 bytes using table layout). Both ISPs actively obfuscate their filtering product by reporting generic 'Server: Apache/2.2.9 (Debian)' or 'Server: Apache' HTTP headers instead of the actual product identifier.

2014-jones-automated §6 detection

All three prior end-to-middle (E2M) schemes — Telex, Cirripede, and Decoy Routing — require an inline flow-blocking component at the participating ISP, which adds latency, introduces a single point of failure, and may violate carrier SLAs. In private discussions with ISPs, the authors found that despite willingness to assist Internet freedom technically and financially, none were willing to deploy existing E2M technologies due to these operational impacts. TapDance removes the inline blocking requirement entirely, requiring only a passive tap and packet-injection capability.

2014-wustrow-tapdance §1, §6 deployment

In 80% of measured paths (72 PlanetLab VPs × 5,000 Alexa targets), at least one intermediate router returns the full IP packet in ICMP time-exceeded replies (RFC1812-compliant), enabling per-hop detection of packet modifications. The majority of these full-ICMP routers reside in the network core rather than the access segment.

2013-detal-revealing §3.2 evaluation

Middleboxes that randomize TCP sequence numbers do not update the sequence numbers inside TCP SACK blocks; tracebox found two PlanetLab VPs with stateful seq-number randomizers that cycled approximately every 20 seconds. When SACK blocks reference sequence numbers outside the current window, the Linux TCP stack waits for a full RTO instead of fast-retransmitting, producing up to 50% throughput degradation in controlled measurements.

2013-detal-revealing §3.3 evaluation

Of 72 PlanetLab vantage points, 7 (~10%) automatically stripped or replaced TCP options (Multipath TCP, MD5, and Window Scale) with NOPs at the very first hop, and 2 VPs always altered TCP sequence numbers. These modifications occurred without any corresponding update to dependent fields, corrupting the TCP stream for higher-layer protocols.

2013-detal-revealing §3.1 evaluation

tracebox can estimate middlebox location with an error of ≤4 hops in 61% of cases; errors above 13 hops (the length of ~60% of paths) are each below 1% individually. Of MSS-modifying middleboxes detected, 52% were located in the network core and only 2.7% close to the source vantage point.

2013-detal-revealing §3.4 evaluation

tracebox identified a transparent HTTP proxy or IDS within a National Research Network (SUNET) that intercepted port-80 SYN probes but not port-21 SYN probes, producing shorter observed path lengths to port 80. It also found proxy misconfigurations causing forwarding loops for non-HTTP traffic, where ICMP replies alternated between two routers indefinitely.

2013-detal-revealing §4.2 evaluation

GFW reassembles both IP fragments and TCP segments for HTTP connections, but its overlap-resolution policy diverges from receiver behavior in documented cases: it prefers the original IP fragment in all overlap configurations except when the challenger is simultaneously left-long and right-long (IP2), and prefers a later left-equal TCP segment over the original (TCP5). The paper tests all 18 possible fragment overlap cases and confirms that placing a banned keyword only in the fragment version GFW discards achieves evasion.

2013-khattak-towards §5, Table 1 (IP2, TCP5) detection

GFW instantiates a TCB upon observing a bare SYN before any SYN-ACK (TCP1), enabling a split-connection evasion: a client sends a low-TTL SYN visible to GFW but not the server, then opens the real connection on the same 5-tuple with a different initial sequence number. GFW tracks the phantom TCB and fails to detect banned keywords on the real, desynchronized connection. This same behavior also renders GFW vulnerable to SYN-flooding-style memory exhaustion.

2013-khattak-towards §5, Table 1 (TCP1) detection

A TTL-limited bare FIN packet (without ACK) is sufficient to induce GFW to tear down its connection state for a live TCP session (TCP6b), because GFW accepts FIN packets that violate RFC 793's requirement for the ACK flag. After induced state teardown, subsequent packets carrying banned keywords on the same connection produce no RST, confirming the monitor has lost track of the flow.

2013-khattak-towards §5, Table 1 (TCP6b) detection

OONI observes that many interception devices deployed in the wild advertise their vendor and model information, making passive device identification feasible from probe-level observations alone. The framework is designed to locate interception devices and then apply probing techniques to fingerprint the specific vendor and product in use.

2012-filast-ooni §2 Goals detection

OONI's traffic manipulation test suite uses bidirectional traceroute comparison: asymmetry between inbound and outbound paths for specific source/destination port pairs is treated as an indicator that traffic is being diverted to an interception device. Additional per-flow indicators include timing differences in packets directed at specific ports and layer-7 header field manipulation detectable at the receiving endpoint.

2012-filast-ooni §5.5 Traffic Manipulation detection

To measure Chinese search engine censorship independently of backbone GET request filtering, the authors split each search engine HTTP GET request across multiple TCP packets so the server would reassemble the full query but routers performing single-packet keyword inspection would not see a complete match. This technique allowed ground-truth measurement of search engine responses free of backbone RST injection interference.

2011-espinoza-automated §3 defense

An adversary aware of a decoy router's location can force decoy-routed flows to be unprocessable by fragmenting all packets below the size of a complete TCP header in the first fragment, preventing flow assignment and forcing the router into expensive reassembly. Alternatively, the adversary can use small-fragment attacks to grow the router's state table, analogous to NAT resource exhaustion. The paper identifies fragmentation-based denial as a harder-to-mitigate attack class than sentinel replay.

2011-karlin-decoy §4.2 detection

TCP flow hijacking by the decoy proxy is practical under an asymmetric routing assumption: expected sequence numbers are recoverable from ACK values in client-originated packets alone, so the decoy router need not observe return traffic. The proxy forges a TCP RST to the decoy destination and mimics its TCP options (timestamp, window scale, SACK) to reduce detectability; these options are conveyed encrypted inside the sentinel's 28-byte TLS random field.

2011-karlin-decoy §3.3 defense

The paper identifies two unresolved fingerprinting surfaces: (1) traffic-shape analysis of packet sizes and inter-arrival times could distinguish Telex flows from normal TLS, and (2) the prototype exhibits detectable deviations from real servers at the IP layer (stale IP ID fields), TCP layer (incorrect congestion windows detectable by early acknowledgements), and TLS layer (different compression methods and cipher-suite extensions). Convincingly mimicking a diverse population of TCP/TLS server implementations is flagged as requiring substantial engineering effort.

2011-wustrow-telex §9 detection

TTL manipulation experiments demonstrated that the GFW injects forged DNS responses at the router level, not at the DNS server: responses to censored domain queries exhibited inconsistent IP ident fields and wildly varying TTL values — consistent with a stateless in-path router — while control (non-censored) responses to the same server showed monotonically increasing ident and stable TTL. The injection was observed exclusively on port 53; identical queries sent to port 80 received no injected responses.

2007-lowe-great §6.4, Table 3 detection

A hybrid two-stage blocking system (IP-redirect first stage, URL-proxy second stage) can be exploited as an oracle to enumerate blocked IP addresses by sending TCP packets with a TTL sufficient to reach the first-stage redirector but insufficient to reach the destination. Non-redirected IPs return ICMP TTL-expired from an intermediate router, while redirected IPs return a SYN/ACK from the web proxy impersonating the destination. A live scan of a /24 subnet confirmed 17 redirected IP addresses, yielding 91 associated hostnames across 9 of those IPs.

2006-clayton-failures §5.2 defense

Brightview's countermeasure requiring a minimum probe TTL of 24 (to prevent low-TTL scans from stopping at the proxy) was bypassed by sending probes with TTL=128 and examining the TTL of returned SYN/ACK packets. The UK web proxy consistently returned TTL=49 (64−15 hops), while Russian destination servers returned TTL=45–49 or TTL=113–238 depending on initial OS TTL settings. The two populations were cleanly distinguishable, defeating the fix with no change to scan logic beyond raising the probe TTL.

2006-clayton-failures Postscript defense

The GFW performs no stateful TCP stream reassembly, inspecting one packet at a time: splitting the blocked keyword '?falun' across two TCP segments is sufficient to evade detection entirely. Cross-device state is also absent — triggering a block on one border AS (e.g., AS9929) had no effect on traffic transiting a different Chinese border AS.

2006-clayton-ignoring §4.1 detection

TCP RSTs are delivered unreliably and different OS stacks apply different validity rules, so a NIDS cannot safely tear down connection state on RST alone; a 'reliable RST' scheme — sending a keep-alive ACK behind every forwarded RST and tearing down state only upon observing a confirming RST from the trusted side — resolves this without violating end-to-end semantics. The cold-start problem (state loss on restart) can be addressed statelessly by stripping payload from unknown-connection packets from untrusted hosts and probing the trusted endpoint with a keep-alive before instantiating state.

2001-handley-network §6.1–6.2 defense

Passive NIDS can be evaded via three fundamental classes of ambiguity: incomplete protocol analysis (none of the four commercial systems tested by Ptacek and Newsham in 1998 correctly reassembled IP fragments), divergent end-system behavior (different OS stacks resolve overlapping TCP retransmissions differently), and topology uncertainty (low-TTL packets may not reach the victim end-system, so the NIDS cannot determine which packets are delivered).

2001-handley-network §1 detection

A traffic normalizer placed inline ('bump in the wire') can eliminate over 70 IP/TCP packet-level ambiguities before a NIDS inspects traffic — including fragment reassembly, TTL restoration, DF flag clearing, IP option removal, and cryptographic IP ID scrambling — leaving the classifier with an unambiguous byte stream and removing the degrees of freedom an attacker needs to evade detection.

2001-handley-network §3, Appendix A defense

The paper proves that any network IDS operating without maintaining complete, OS-specific per-connection state cannot reliably reconstruct the byte stream seen by the end-system. TCP and IP reassembly ambiguities guarantee unavoidable blind spots unless the IDS performs full per-target OS emulation—a fundamental architectural limitation, not an implementation bug, that applies equally to any DPI-based censor.

1998-ptacek-insertion §5 defense

IP-level fragment overlap attacks operate independently of TCP: crafting overlapping IP fragments whose reassembly by the IDS yields benign content while the end-system's reassembly yields the true payload. The paper demonstrates this is a separate attack surface from TCP-level evasion, exploitable below the transport layer before any TCP stream reconstruction begins.

1998-ptacek-insertion §3 defense

Different operating systems apply different precedence rules when TCP segments overlap—some implementations use 'first data wins,' others 'last data wins.' An IDS applying a single universal reassembly policy will systematically diverge from the actual target end-system whenever overlapping segments appear, creating a predictable and repeatable evasion surface that is an inherent consequence of policy misalignment rather than a configuration flaw.

1998-ptacek-insertion §4 defense