← all systems · see it in The System
Denial-of-service attack
A cyberattack in which the perpetrator seeks to make a machine or network resource unavailable to its intended users by exhausting a finite resource it depends on (src:wp:denial-of-service-attack; RFC 4732). Walked as the DoS measurement and protocol literature describes it: the finite resources a networked service runs on (link bandwidth, connection state, CPU, memory, application sessions); the attack sources (single hosts, botnets, reflectors and amplifiers, volunteers, hijacked browsers); the kinds of attack traffic (volumetric floods, reflection and amplification, protocol-state exhaustion, application-layer, low-and-slow, algorithmic-complexity, low-rate shrew, crash and permanent attacks, telephony DoS); the exhaustion at the bottleneck and the denial to legitimate users; the attack's command; the booter market and extortion campaigns; the motives; and the defenders it meets (filtering, SYN cookies, blackholing, scrubbing, police takedowns). History from Saltzer and Schroeder's 1975 'denial of use' to the 22.2 Tbit/s record of 2025.
Properties published scienceour proposal
Sourced fields, and our own readings.
| born at | 1975 (named); 1996 (first publicised SYN-flood attacks); 1999 (first documented botnet DDoS) |
| born precision | year |
| boundary | Inside: the attack while it runs: the attacker's command (target, vector, start and duration), the sources sending on its behalf while they send (bots, reflectors, volunteers' machines, hijacked browsers, rented servers), the attack traffic itself, and the exhaustion it causes at the target's bottleneck. Outside: the target and its legitimate users (they are what is denied), the Internet and its protocols that carry the traffic, the botnet as a standing infrastructure (its own walk), the booter business and the operators as people (threat actor), and the defenders (filtering, scrubbing, blackholing, police). A reflector belongs to the attack only for the packets it is tricked into sending; its operator is outside. |
| nested level | Run by threat actors over the Internet; built from attack sources and attack traffic; a kind of cyberattack (Saltzer and Schroeder's third class of violation, unauthorized denial of use); sold as a service in the cybercrime economy; part of cyberwarfare when states use it. |
| scale | technological |
Measures (37) published science
| characteristic clock: TCP half-open state reclamation timer (example) | 75 s |
| characteristic clock: TCP minimum retransmission timeout (shrew period) | 1 s |
| characteristic clock: attack duration, share under 10 minutes (2001-2004) | 60 % |
| average delay to blackhole an attack at a major IXP | 1.16 minutes |
| market suppression after the HackForums booter ban | 13 weeks |
| time for a seized booter to reappear on a new domain | 3 days |
| NTP amplifier reduction after notification campaign | 92 % over 13 weeks |
| half-open connection memory (Linux 2.6.10 sock structure) | >1300 bytes |
| typical default TCP backlog | a half-dozen to several dozen connections |
| bandwidth amplification factor, worst case (NTP monlist) | 4670 x |
| amplification factor, memcached | up to 50000 x |
| amplification factor, NTP (US-CERT table) | 556.9 x |
| protocols found susceptible to amplification | 14 protocols |
| ASes found lacking egress filtering | 2692 ASes |
| attacks seen by backscatter, 2001-2004 | 68000 attacks |
| attacks seen 2015-2017 (telescope plus honeypots) | 20000000 attacks |
| Web sites in .com/.net/.org involved with attacks daily | 3 % |
| targets migrating to a DDoS protection service after attack | 4.3 % |
| amplification attacks in one day at one IXP | 2608 attacks/day |
| share of IXP attacks blackholed | 3.82 % |
| longest attack seen at the IXP (SSDP) | 7.5 days |
| attacks on one cloud provider in 2021 | 3000000 attacks |
| SYN-flood share of attacks at OVHcloud, 2021 | 45 % |
| booter price for a month of attacks | 10-20 USD |
| booter VIP vs basic package (booter B) | 178.84 vs 19.83 USD |
| delivered vs advertised booter VIP rate | 25 % |
| reflectors per booter attack | 100-1000 reflectors |
| attack reduction after FBI Xmas2018 operation | 27-37 % for 10 weeks |
| attacks on the Netherlands after Webstresser takedown | +146 % |
| DDoS mitigation market, 2018 | 1.94 billion USD |
| record attack bandwidth (Cloudflare claim, 23 September 2025, 40 seconds) | 22.2 Tbit/s |
| record HTTP request rate (Google, October 2023) | 398 million requests/s |
| shrew throughput cut (WAN, 909 kb/s attack) | 9.8 to 1.2 Mb/s |
| IDS traffic dropped after six minutes of algorithmic attack | 71 % |
| cost to break Tor's directory protocol by DDoS | 53.28 USD per month |
| phones needed in simulation to block a US state's 911 for days | <6000 bots |
| first documented botnet DDoS (Trin00, University of Minnesota) | 1999 year |
Open contentions (4) open contention
Where sources or readings disagree. Both stay on the map.
- C-1651 What was the first denial-of-service attack?
booked: The map keeps three firsts: the class named in 1975, the first widely seen attack in 1996, and the first documented botnet attack in 1999.
the other reading: Panix 1996 as 'thought to be the first DoS attack' (encyclopedia); Trin00 1999 as the first documented case (Osterweil et al.: earlier attacks exist 'in anecdotes and recollections'); Saltzer and Schroeder 1975 as the first naming of the class; telephone line jamming before the Internet (1981 song-number flood, an unintentional case) - C-1652 Is a denial-of-service attack a crime or a form of protest?
booked: The map records it as a crime in most countries and keeps the claim that it can be legitimate protest.
the other reading: criminal attack; a form of protest (the 2013 petition's claim) - C-1653 Are defenders winning or losing against denial-of-service attacks?
booked: The map rules neither: one study says defenders are losing, another finds the internet's core has room to spare.
the other reading: defenders losing (architectural asymmetry, centralised scrubbing against distributed attacks); defenders holding at the core (spare capacity, automated blackholing within minutes); a stalemate that reads differently at the edge and the core - C-1654 Did the 2018 FBI takedown of attack-for-hire sites reduce attacks?
booked: The map keeps both: recorded attacks fell by about a third for ten weeks, while traffic reaching victims did not measurably change.
the other reading: Collier et al. 2019: a third fewer attacks for at least 10 weeks, lasting change in market structure; Kopp et al. 2019: reduced traffic to reflectors, no significant effect on victims; a booter back in three days
Cross-scale patterns seen here (19) our proposal
Shapes that recur when the same kind of problem is solved at different scales, and how each shows in Denial-of-service attack. Each links to every other system that shows it.
- same role, different builds (structural)
One job, denying a service to its users, done by unrelated builds: filling a link (floods), filling a backlog (SYN), holding connections (slow POST), driving a hash table to its worst case, crashing the kernel with one packet, flashing bad firmware, occupying phone lines. In nature a tiger moth deni - same shape, opposite sign (structural)
One process, holding half-open TCP connection state for an unverified client, run in both directions: the SYN flood forces the server to hold state for clients that will never answer; SYN cookies push the state into the packet the client must return, so the server holds none until the client proves - templating (structural)
Attack code reused unchanged across attackers: the 1996 Phrack SYN-flood code 'widely distributed and used', the Panix sample code reused against Sprint, EarthLink and E-Trade the next year; booter method menus offer the same named vectors (DNS, NTP, SNMP, LDAP) across services. - The carry (structural)
A place that holds only so much: the backlog of half-open connections (a half-dozen to several dozen) and its reclamation timer (75 s in one example) set a settling time; the SYN flood is sized to the capacity and paced to the wait. The bang.c variant shows overflow passed on: once the link overflow - The Payment to Entropy (structural)
Each hand-off of a flood pays a toll: the Internet's capacity attenuates an attack's throw weight, and the farther the target from the sources the less arrives across congested links; a premium booter delivered about 25% of the rate it advertised; past saturation a host's useful work falls (livelock - trade one property for another (structural)
Every DoS defence pays with another property: SYN cookies buy stateless setup and pay in TCP options; a lower connection limit protects the machine and makes the attacker's job easier, a higher one risks crash or livelock; puzzles stop mass senders and penalise slow legitimate devices; blackholing s - Umwelt: each system carves its own slice of one shared medium (structural)
The same packets read differently at each layer: layering hides DDoS semantics from the network and transport layers, so a router sees addresses and ports, a scrubbing centre sees application semantics, and an attack built in the slice the defender cannot read (encrypted application traffic, legitim - who opens the door (structural)
The receiver opens the door ahead of time: an unbound LISTEN accepts any client not known in advance (installed by the server's application and OS); an open resolver or NTP server answers anyone (installed by its operator). The attack walks in through ports the receivers built. - causal ceiling (partial)
A limit on what any defender can know: 'in principle it is not possible to distinguish between a sufficiently subtle DoS attack and a flash crowd'; and disruption on congested transit links is 'unobserved (and unobservable)'. Partial: the RFC states it in principle, but the limit is informational (i - coordination-before-mapping (partial)
MyDoom's infected hosts flooded one target together at a hard-coded date and time, with no further interaction and no command channel: coordination by clock alone. - lightning (partial)
The lightning test: the same denial happens with no attacker, met trivially by a surge, a flash crowd (Slashdot, Michael Jackson's death in 2009) or a misconfigured router flooding NTP servers. The deliberate attack is a commanded, sized and timed discharge (a 40-second 22.2 Tbit/s burst), so the at - namespace completeness (the map predicts its gaps) (partial)
A survey of UDP protocols whose replies exceed their requests (14 protocols, 2014) predicts which services will be abused before the abuse is seen: attackers 'already started' on protocols beyond DNS, and newly found vectors (ARMS, OpenVPN, Ubiquiti discovery) appeared in IXP traffic in 2020-21. - red-queen (partial)
Both sides keep improving: amplifiers are cleaned (NTP -92%), new vectors appear (+500%); booters are seized, return on new domains; mitigation grows to a $1.94 billion market while attacks grow 'faster than network growth'. Whether neither gains ground is contested: Osterweil et al. say defenders a - runaway from an unstable balance (partial)
Saturation is a balance point: below it more load means more work; past it the system can run away (livelock: more load, less useful work); bang.c's lost RSTs feed more retransmissions once the link overflows. - same toolkit, independent builds (partial)
Separate booter services build their attacks from the same shared pool of public reflectors and the same protocols; seizing one front end leaves the reflectors 'utilized by third parties'. Partial: whether the booters are independent builds or copies is not shown. - symmetry breaking (name the cut) (partial)
Before: the network treats every packet alike and a subtle attack cannot be told from a crowd. Defence must break that symmetry at a named cut: 'find something that legitimate users can do but attackers can't' (puzzles, reverse Turing tests, reachability), each cut with its own cost to legitimate us - holding / entrainment (surface)
The shrew attack sets the clock of every TCP flow on the link: bursts at the RTO period keep them timing out together. Surface: the attacker is outside, not a whole holding its parts. - seed (surface)
The 1996 publication of SYN-flood code, and the Panix sample code, were followed by attacks on many companies the next year: a small release followed by a spread. Surface: a trigger that spread a technique, not combinatorics unpacked. - the whole sets the local reference (surface)
Detection reads 'abnormal' against the service's own peace-time profile (provisioned load, key completion indicators: the usual path of value through the application). Surface: a baseline, not a whole whose local readings fail to close.
Valence: what it trades, and with whom (17) published science
Each exchange with a neighbour, read from both sides: who gains, who pays, and what carries it.
- Human (Homo sapiens) via sys:human · Denial-of-service attack gain
- Internet via sys:internet · Denial-of-service attack gain
- Internet protocol suite (TCP/IP): the layered datagram protocols every host and router runs, the RFC record that specifies them, and the community that keeps them via seed:tcp-ip · Denial-of-service attack gain
- Operating system via sys:operating-system · Denial-of-service attack gain
- Police via sys:human-society · Denial-of-service attack loss
- Malware via sys:computer · Denial-of-service attack gain
- Intrusion detection system via sys:internet · Denial-of-service attack gain
- Threat actor via sys:threat-actor · Denial-of-service attack gain
- Botnet via sys:internet · Denial-of-service attack gain
- Organization via sys:human-society · Denial-of-service attack gain
- Firewall (computing) via sys:internet · Denial-of-service attack mixed
- Computer network via sys:internet · Denial-of-service attack gain
- Reflectors and amplifiers (open DNS resolvers, NTP, memcached, CLDAP, SSDP servers) via sys:internet · Denial-of-service attack gain
- DDoS protection service (scrubbing centres) via sys:internet · Denial-of-service attack loss
- Network telescope (backscatter observers) via sys:internet · Denial-of-service attack loss
- Cyberwarfare via sys:internet · Denial-of-service attack gain
- Cybercrime and its underground economy via sys:cybercrime · Denial-of-service attack gain
Partners and parts published science
Built from / builds (9)
- Threat actor
- Botnet
- Reflectors and amplifiers (open DNS resolvers, NTP, memcached, CLDAP, SSDP servers)
- Attack traffic (packets, requests, calls)
- Booter / stresser service (DoS-as-a-service)
- Attackers against defenders for control of computing
- Cyberwarfare
- A botnet's flood against a target's capacity
- Attack traffic against legitimate users for the bottleneck
Sits inside / holds (3)
expressed by (16)
- Templating / Standardization
- Amplification/Scale-Up (Manufacture & Distribution)
- Sensing/Scanning
- Evade Detection
- Encirclement
- Precision Attack
- Ranged Attack
- Neutralize Defenses
- Resource Denial/Attrition
- Subversion & Co-optation
- Prepositioning & Latency
- Leverage Host/Amplification
- Divide & Isolate
- Cascade Induction
- Resilience & Repair
- Redundancy & Diversity
floor filled by (17)
- Human (Homo sapiens)
- Internet
- Internet protocol suite (TCP/IP): the layered datagram protocols every host and router runs, the RFC record that specifies them, and the community that keeps them
- Computer
- Operating system
- Malware
- Threat actor
- Botnet
- Telephone network (circuit-switched carrier)
- Command and control (C&C)
- Cybercrime and its underground economy
- Great Cannon (backbone-injected ephemeral botnet)
- Reflectors and amplifiers (open DNS resolvers, NTP, memcached, CLDAP, SSDP servers)
- Attack traffic (packets, requests, calls)
- Booter / stresser service (DoS-as-a-service)
- DDoS protection service (scrubbing centres)
- Cyberwarfare
Kinds (19)
- Threat actor
- SYN flood
- Distributed denial-of-service (DDoS) attack
- Volumetric flood
- Reflection and amplification attack (DRDoS)
- Application-layer (layer 7) DDoS
- Low-and-slow attack (Slowloris, slow POST, RUDY, slow read)
- Algorithmic complexity attack
- Low-rate TCP-targeted (shrew) attack
- Crash attack (ping of death, teardrop, nuke)
- Permanent DoS (phlashing, bricking)
- Telephony DoS (TDoS)
- Degradation-of-service (pulsing zombies)
- Yo-yo attack
- Unintentional DoS (flash crowd, Slashdot effect)
- Advanced persistent DoS (APDoS)
- Triggered lockout / quota exhaustion
- Physical DoS
- Legal DoS
Notes published science
- A burglar who never touches the safe: in 2003 two researchers sent a security monitor less traffic than a dial-up modem carries, each packet chosen so its hash table degenerated into a list, and after six minutes it was dropping 71% of what it was meant to watch. (src:crosby-2003-algorithmic)
- The shrew and the moth play the same trick: a DoS attack can win with tiny bursts timed to TCP's one-second retry timer, cutting a flow from 9.8 to 1.2 Mb/s; tiger moths jam bat sonar the same way, crowding clicks into the moment the bat is listening for its echo, and bats fight back by lengthening their buzz. (src:kuzmanovic-2003-shrew; src:fernandez-2022-moth-jamming)
- Busting the shop doesn't empty the warehouse: when the FBI seized 15 booter sites in December 2018, traffic to the reflectors dropped but traffic hitting victims did not, because the open servers that amplify attacks stayed online for anyone to use; one seized booter was back on a spare domain, registered six months earlier, within three days. (src:kopp-2019-booter-takedown)
Layers
The ways it takes in its world, where they meet, and how it acts back, each from L1 (where it meets the world) up.
the boardroom (shared floors)
- L4: The bottleneck at the target: the finite resource (link, backlog of half-open connections, CPU, connection pool, phone lines) where attack traffic, legitimate traffic and defences meet; whichever fills it decides who is served
- L5: The attack command: target, vector, start and duration, given by the attacker's own client, Stacheldraht's handlers (up to a thousand agents each), a botnet's command and control, a booter's panel, or a date hard-coded in malware (MyDoom)
- L6: Campaigns and the market: booters and their customers, extortion schemes, diversions and weeks-long campaigns, and the protection market that answers them ($1.94 billion in 2018)
- L7: The purpose: extortion, rivalry between game servers and shops, protest, censorship, revenge and vendettas, war, or nothing beyond being able to; carried by the attacker (runsOn threat actor)
the denial-of-service attack as DoS researchers describe it: the finite resources a service runs on, up to the attacker's purpose (reader's view)
- L1: The finite resources a networked service runs on: link bandwidth, router and server CPU, memory, connection tables, disk, application sessions, phone lines
- L2: Attack sources: a single well-connected host, botnets of thousands, reflectors and amplifiers, volunteers' machines, bystanders' browsers
- L3: Attack traffic by vector: volumetric floods (UDP, SYN, ICMP), reflection and amplification with spoofed sources, application-layer requests, low-and-slow connections, algorithmic-complexity inputs, low-rate bursts, malformed packets, automated calls
- L4: Exhaustion at the bottleneck: the link saturates, the backlog fills, the CPU livelocks, the connection pool is held; legitimate requests are crowded out and the service is denied
- L5: The attack command: target, vector, start, duration; client and handlers, botnet command and control, booter panels, a hard-coded date
- L6: Campaigns and the market: booter services at tens of dollars a month, extortion schemes, diversions, multi-week campaigns; the protection market answering them
- L7: The purpose: extortion, rivalry, protest, censorship, revenge, war, or none beyond being able to
way in: Attack sources: the machines that can send
- L1: Internet-connected machines with bandwidth: PCs, IoT devices, servers, browsers; attack capacity easier to gain than defensive capacity
- L2: Enlisted: infected with malware (botnets of thousands of hosts), volunteered (LOIC in Operation Payback), or hijacked in passing (the Great Cannon programming bystanders' browsers)
- L3: Each source sends: TCP, UDP or ICMP packets, HTTP requests; the Cannon injected its script in about 1.75% of requests
way in: Customers' orders (booter market)
- L1: Booter websites with simple web front-ends; payment by PayPal or cryptocurrency
- L2: Packages: $10-20 a month typical; a VIP package at $178.84 against $19.83, promising 80-100 Gbps
- L3: The order: target, method and duration (researchers' self-attacks configured to last 5 minutes)
way in: Defenders' responses
- L1: The defended path: the target's stack, its ISP and IXP, DDoS protection services and their scrubbing centres
- L2: Detection: the service notices it is under attack; detection and remediation often in different places by different providers
- L3: Response: blackholing (3.82% of IXP attacks, 1.16 min average delay), scrubbing, firewall rules, rate limits, SYN cookies, universal hashing
way in: Law and police action
- L1: Laws: 156 countries (80%) with cybercrime laws; the UK's 10-year maximum for DoS
- L2: Actions: arrests, domain seizures, police visits and warnings, search adverts aimed at young users
- L3: Effect on supply: seven booters offline at once in December 2018; one back on a new domain three days later
way in: Legitimate load and flash crowds
- L1: Legitimate users' requests to the same service
- L2: Peaks: flash crowds (the Slashdot effect, Michael Jackson's death in 2009), evening peaks 5-11 pm; in principle indistinguishable from a subtle attack
- L3: The cover: pulsing floods and disguised application requests look like heavy legitimate load
way in: New attack vectors from protocol design
- L1: Protocols whose responses exceed their requests, or which hold state for unverified clients
- L2: A vector found: 14 protocols in 2014; OpenVPN, ARMS, Ubiquiti discovery rising 500% in 2020-21; HTTP/2 Rapid Reset in October 2023
- L3: Added to tools: booter method menus name DNS, NTP, SNMP, LDAP, PORTMAP
way in: Permission to forge the source address
- L1: Edge routers that forward packets whatever their source address
- L2: No source address validation: up to 2,692 ASes without egress filtering; deploying it only helps other networks
- L3: Forged-source packets carried to reflectors and victims; the largest attacks all rely on spoofing
way in: Reflectors and amplifiers
- L1: Public UDP services: open DNS resolvers, NTP, SNMP, memcached, CLDAP, SSDP, game and P2P protocols
- L2: Amplifiers: millions of public hosts; a notification campaign cut NTP amplifiers by more than 92%, yet exploited protocols stay a threat for decades
- L3: Amplified replies to the spoofed victim: up to 4670 times (NTP), up to 50,000 (memcached); booter attacks use about 100-1,000 reflectors
way in: The target's own protocol settings
- L1: The target's protocol implementation: TCP's unbound LISTEN, the backlog, reclamation timers, the retransmission timeout, hash tables
- L2: The parameter: backlog of a half-dozen to several dozen; a 75 s reclamation timer (511 s in 4.4BSD-Lite); minimum RTO of 1 s; predictable hash functions
- L3: The attack tuned to it: a barrage no larger than the backlog, renewed as state is reclaimed; bursts at the RTO period; inputs that all collide
way in: World events
- L1: World events: wars, summits, holidays
- L2: The trigger: Russia's invasion of Ukraine (February 2022); Zelensky at Davos (January 2024); school holidays
- L3: Targets chosen to match: Ukrainian government and finance, then UK, US and German allies; Swiss federal websites
way out: Application-layer and low-and-slow
- L1: The server software unable to serve
- L3: Requests that over-exercise a function; bodies sent at 1 byte per 110 s; inputs that collide in hash tables
- L4: Connection pool, sessions or CPU held while bandwidth stays free
- L5: Function or connection type chosen
way out: Cascade to what depends on the target
- L1: Networks unreachable; sites behind Dyn unreachable for hours; the Tor network brought down
- L3: Routing churn trips route flap damping; directory votes fail
- L5: A shared dependency targeted: a DNS provider, routers, Tor's directory authorities
way out: Collateral damage and backscatter
- L1: Other services degraded; backscatter lands across the address space, where network telescopes sample it
- L3: Replies and attack traffic share congested links
- L4: The victim answers each spoofed packet
way out: Crash
- L1: The operating system crashes (WinNuke's blue screen)
- L3: One malformed packet: a fragmented ICMP echo over 65,535 bytes, overlapping fragments, out-of-band data to port 139
- L5: A known crash bug chosen
way out: Extortion message
- L1: An e-mail read by the target's staff
- L5: The demand composed
- L6: Low-level attack then a warning: pay in bitcoin or a larger attack follows
way out: Permanent DoS
- L1: The device bricked until repaired or replaced (BrickerBot)
- L3: Corrupt firmware flashed
- L5: Device chosen through its remote management flaw
way out: Protocol-state exhaustion (SYN flood)
- L1: Server memory: over 1300 bytes per half-open connection (Linux 2.6.10)
- L3: SYNs from unresponsive spoofed addresses
- L4: Backlog of half-open connections kept full
- L5: Barrage size, frequency and spoofed addresses chosen
way out: Telephony flood
- L1: The victim cannot make or receive calls, emergency calls included; in simulation under 6,000 phones block a state's 911 for days
- L3: Repeated automated calls; spoofed caller ID
- L4: Lines occupied
- L5: Number chosen
way out: Volumetric flood
- L1: The target's links saturated: terabit attacks on Krebs and Dyn in 2016; 22.2 Tbit/s for 40 seconds in September 2025
- L3: Sources and reflectors send; a VIP booter delivered about 25% of the advertised rate; traffic thins with distance across congested links
- L5: Attack started for a set duration