A denial-of-service attack (DoS attack) is a deliberate attempt to prevent legitimate users from accessing a computing service, network, or other digital resource, or to degrade that resource below an acceptable level of operation. It targets availability, one of the principal objectives of information security, rather than necessarily seeking to steal information or modify data. A distributed denial-of-service attack (DDoS attack) uses multiple sources to produce the disruptive effect. Distribution can increase aggregate traffic, complicate filtering, and make the attack less dependent on any single participating machine.
Denial of service can result from overwhelming an Internet connection, filling a server’s connection tables, consuming application resources, or triggering a software defect that causes a crash or persistent failure. Some attacks require enormous traffic volumes; others exploit an asymmetry in which a small, inexpensive input causes disproportionately expensive processing. Their significance extends beyond individual websites because shared naming systems, hosting platforms, communications networks, and other dependencies can transmit disruption to many unrelated organizations.
The subject therefore spans network protocols, software design, distributed computing, incident response, and the economics of shared infrastructure. Understanding it requires distinguishing the mechanism of an attack from its visible symptoms, the resources exhausted from the machines sending traffic, and reported attack size from the actual effect on users.
Availability, resource exhaustion, and attack classification
Availability is not simply whether a machine responds. A service may remain reachable while taking minutes to answer, rejecting most transactions, returning incomplete results, or becoming inaccessible from particular networks. Denial of service can therefore be total, intermittent, geographically limited, or restricted to a function such as account login or payment processing. Its practical severity depends on the service’s purpose and the duration and distribution of the impairment.
The defining feature is deliberate interference. An unexpected surge of legitimate demand, often called a flash crowd, can produce nearly identical symptoms without being an attack. An accidental software failure can also deny service. Conversely, attack traffic may consist entirely of syntactically valid requests. Neither a high traffic rate nor a protocol violation is necessary to establish malicious denial of service; intent, context, traffic behavior, and system effects must be considered together.
A useful model treats each limiting resource separately. For resource , let represent the arriving workload and its sustainable processing capacity. In a simplified steady-state queueing model, utilization is
As utilization approaches one, waiting times can rise sharply; if incoming work persistently exceeds processing capacity, queues grow until requests are dropped or another limit is reached. Real systems have bursts, multiple request classes, finite buffers, and adaptive controls, so this model is explanatory rather than a complete prediction. It nevertheless shows why a server can become unusable before every processor is fully occupied.
Attacks are commonly divided into volumetric, protocol or state-exhaustion, and application-layer categories. Volumetric attacks consume transmission capacity. State-exhaustion attacks overwhelm structures that track connections, sessions, or protocol exchanges. Application-layer attacks target the work performed by software, such as searching a database, generating a report, or authenticating a user. These categories overlap: an application attack can also generate substantial bandwidth, while a network flood can force expensive processing in security equipment.
Not all denial of service is a flood. A carefully formed input may exploit a buffer overflow, an unhandled exception, or an infinite loop and terminate a process with very little traffic. A local user may exhaust disk space, process identifiers, or file descriptors. Algorithmic-complexity attacks exploit worst-case behavior in an algorithm, making input size a poor predictor of processing cost. The broad concept includes all these mechanisms, although public reporting often uses “DDoS” more narrowly for coordinated network or application flooding.
Network flooding and protocol-state exhaustion
Network-layer denial of service exploits limits in moving and processing packets. A link has a finite bit rate, but forwarding equipment also has a finite packet-processing rate. Large packets may exhaust bandwidth, whereas small packets can impose a high per-packet workload without reaching the same bit rate. A router, network interface, firewall, or load balancer can therefore fail at a point different from the server it protects. Capacity measured only in gigabits per second does not describe all relevant limits.
Traffic floods using the User Datagram Protocol, or UDP, are prominent because the protocol does not require an established connection before a datagram is transmitted. Depending on the target, incoming packets may consume link capacity, trigger application work, or cause the host to perform lookup and error-handling operations. Internet Control Message Protocol traffic can likewise be used to consume bandwidth or processing resources. The effect depends on packet handling, filtering, and the location of the bottleneck, not merely the protocol name.
The classic SYN flood targets connection establishment in the Transmission Control Protocol (TCP). A normal TCP connection begins with a three-part handshake. A server receiving the initial synchronization request may retain information while awaiting completion. A sufficiently large number of incomplete handshakes can exhaust the capacity reserved for pending connections, preventing legitimate clients from establishing sessions. Forged source addresses can make replies go elsewhere, but spoofing is not essential to every SYN-flood variant.
SYN cookies reduce this particular state burden by encoding information about a pending connection in the server’s initial sequence number. The server can validate a later response without retaining the usual per-connection record during the initial exchange. This mechanism addresses a specific allocation problem; it does not create additional upstream bandwidth, eliminate packet-processing costs, or prevent exhaustion after connections become established. Practical implementations also involve protocol-option and performance considerations.
State exhaustion can occur in intermediaries even when the destination server has adequate resources. Stateful firewalls and network address translators maintain connection-tracking entries. Load balancers hold flow mappings, and encrypted-service front ends may retain handshake state. A large population of short-lived or incomplete flows can fill these structures. The observable symptom may be widespread connection failure despite low application-level CPU use.
Historical malformed-packet attacks illustrate a different network mechanism. The “ping of death” exploited improper handling of oversized reassembled Internet Protocol packets, while “teardrop” attacks exploited defects in fragment reassembly. Their effects depended on vulnerable implementations rather than an inherent requirement that all systems fail on such traffic. Patching the defect could remove the failure mode, unlike volumetric flooding, where any finite link can still be overwhelmed by sufficient traffic.
Reflection, amplification, and address spoofing
A reflection attack sends requests to third-party systems while making those requests appear to originate from the intended victim. The third parties then send their replies to the victim. Reflection separates the apparent response sources from the attacker’s own infrastructure and can aggregate traffic from many independently operated servers. The reflectors need not be compromised: a legitimate service may become an unwitting participant if it accepts a request with a forged source address and sends a response to that address.
Amplification occurs when the response is larger, or otherwise more burdensome, than the initiating request. A simplified bandwidth amplification factor is
where both quantities must use consistent accounting conventions. Measurements that include only application payload are not directly comparable with measurements that include transport, network, and link overhead. Some mechanisms produce several response packets per request, so packet-rate amplification can differ from byte-rate amplification. A high theoretical factor also does not guarantee a large attack: accessible reflector capacity, rate limits, routing, and congestion constrain the result.
The Domain Name System (DNS) has been used for reflection because some queries can elicit responses substantially larger than the query. Exposed recursive resolvers have been a particular concern, although authoritative DNS infrastructure can also be involved. Other historically abused services include the Network Time Protocol and connectionless discovery protocols. In 2018, publicly reachable memcached services were exploited for unusually large UDP amplification attacks. The common property is a request-response behavior that can be induced without first verifying that the apparent source is reachable by the requester.
IP address spoofing is the principal enabling condition for conventional UDP reflection. Unlike a completed TCP handshake, a single UDP request does not inherently establish that the sender can receive traffic at the claimed source address. Source-address validation at network boundaries can reject packets whose claimed origin is inconsistent with the network from which they arrive. The Internet Engineering Task Force documented ingress filtering in RFC 2267 in 1998 and its successor, RFC 2827, in 2000; the latter is associated with Best Current Practice 38.
Deployment is complicated by asymmetric routing, multihomed networks, and the need to distinguish valid paths from forged origins without discarding legitimate traffic. Validation mechanisms must therefore match the operator’s routing environment. Restricting service exposure, limiting response rates, and avoiding unnecessarily large unauthenticated responses address other parts of the reflection chain.
Reflection and amplification are related but distinct. An attack can reflect traffic without substantially amplifying it, and an attacker-controlled application can generate amplification without third-party reflection. This distinction matters when locating responsibility and choosing defenses: reducing response size does not stop source spoofing, and securing one class of reflector does not remove other reflection services.
Application-layer attacks and asymmetric computation
Application-layer denial of service targets the operations that a service performs on behalf of clients. Requests may look normal to network devices because they use valid HTTP messages, complete transport handshakes, and interact with public functions. The relevant limit may be database connections, worker threads, memory, storage input/output, or access to a downstream service. An application attack can therefore be effective at a traffic rate far below the capacity of the network connection.
The central issue is often unequal cost. A short request can trigger a complex search, a large document conversion, an expensive password-verification operation, or repeated access to uncached data. If an attacker can choose requests that maximize server work while minimizing client cost, adding network bandwidth does little to remove the bottleneck. Computational complexity becomes directly relevant when particular inputs cause unusually expensive parsing, pattern matching, or data-structure operations.
Low-rate attacks exploit resource retention rather than rapid request completion. A server may reserve a connection, buffer, or execution context while waiting for a request to finish. Carefully paced traffic can keep many such resources occupied. The Slowloris technique, publicly described in 2009, became a well-known example involving incomplete HTTP request headers. Its effectiveness depends strongly on server architecture, timeout behavior, and how incomplete requests are managed; it is not a universal property of HTTP services.
Protocol multiplexing changes the economics of application attacks. HTTP/2 permits multiple streams within a connection, improving ordinary efficiency but creating new interactions between stream creation, cancellation, and server work. In October 2023, major service providers disclosed the “Rapid Reset” attack associated with CVE-2023-44487. Repeated stream creation and cancellation allowed attackers to impose substantial processing work. Google reported an attack peaking at approximately 398 million requests per second. The episode showed that limits on simultaneous streams alone need not adequately bound the rate at which streams are created and reset.
Transport Layer Security protects confidentiality and integrity but does not make a service immune to denial of service. Cryptographic handshakes and connection management consume resources, while encryption can restrict what upstream devices can inspect. Conversely, terminating encrypted connections at a sufficiently provisioned front end can isolate origin servers from some costs. The placement of termination and the allocation of processing are therefore important architectural choices.
Application defenses must understand semantics. A login request, a cached image retrieval, and a report-generation request should not necessarily consume the same quota. Authentication can support differentiated treatment, but authenticated users or compromised accounts can still generate abusive workloads. Likewise, a valid API key establishes an identity or entitlement, not that every authorized-looking request is harmless.
Botnets, attack services, and operational organization
Distributed attacks commonly rely on a botnet: a collection of devices controlled through unauthorized software or access. The individual devices, often called bots, generate traffic or perform application requests under coordinated direction. Botnets have included personal computers, servers, routers, cameras, and other Internet-connected equipment. Their usefulness comes not only from aggregate bandwidth but also from the diversity of network locations and addresses, which complicates broad blocking.
Botnet control and attack traffic are separate functions. A command channel can distribute targets and timing while the participating devices send traffic directly. Some botnets use centralized control, while others use peer-to-peer mechanisms or multiple fallback channels. Removing one controller may disrupt a particular organization but does not necessarily remove the malicious software from infected devices. Conversely, disinfecting devices and changing exposed credentials can reduce attack capacity even when controllers remain active.
The Mirai botnet, first widely observed in 2016, demonstrated the scale possible when large populations of embedded devices were reachable with weak or default credentials. Its software targeted devices such as cameras and digital video recorders. Publication of its source code contributed to the appearance of related variants. Mirai is significant not because embedded devices were the first distributed attack sources, but because it showed how insecure consumer and infrastructure-adjacent equipment could be aggregated into substantial Internet attack capacity.
Not all distributed sources are compromised. Attackers can rent computing resources, abuse trial accounts, operate dedicated infrastructure, or use proxy networks. Browser-based participation and explicitly coordinated voluntary traffic have also appeared historically. A large number of source addresses is therefore not sufficient evidence that an incident involves a traditional malware botnet. Malware analysis, provider records, and network observations may be needed to distinguish different sources of coordination.
Commercialized attack services, often advertised as “booters” or “stressers,” lower the organizational barrier by selling access to disruption capability. Stress testing can be legitimate when the relevant infrastructure owners have authorized it; branding alone does not establish that authorization. These services may draw on botnets, reflection infrastructure, or rented servers, and their advertised capacity may not correspond to independently measured performance.
Motives include extortion, competitive disruption, harassment, ideological protest, and interference with online gaming. Threats to launch an attack may be used to demand payment even when the sender has limited capability. Denial-of-service extortion is distinct from ransomware, which generally involves malicious control of data or systems, although the techniques can occur together. Claims of responsibility and political statements are evidence about what a claimant says, not by themselves reliable proof of who operated the attack.
Historical development and major incidents
Denial-of-service problems predate the commercial Internet, arising wherever shared computing resources could be exhausted or disrupted. Their modern networked form became especially visible during the 1990s as public services expanded and protocol implementations encountered hostile traffic at scale. The development of distributed attack tools subsequently changed the problem from controlling a single source to coordinating responses across many networks.
In September 1996, the New York Internet service provider Panix suffered a widely documented SYN-flood attack. The incident helped make TCP connection-state exhaustion a prominent operational security issue. Distributed attack tools such as Trinoo and Tribe Flood Network appeared in the late 1990s, combining compromised machines with coordinated flood generation. These developments established patterns that remain recognizable: separate control and traffic-generating components, multiple flooding methods, and the use of third-party systems.
In February 2000, attacks disrupted major commercial websites, including Yahoo, eBay, and Amazon. The incidents attracted broad attention because they demonstrated that highly visible Internet businesses could be made unavailable without an attacker first taking control of the businesses’ application data. They also encouraged greater coordination among service providers, incident-response teams, and law-enforcement agencies. The consequences were not limited to the duration of individual outages; they influenced how online organizations treated availability as a security concern.
In 2007, a series of attacks affected websites and services in Estonia, including government, media, and banking systems. The episode became an important case in discussions of national dependence on digital services. It also illustrated the difficulty of attribution: traffic origin, political context, public encouragement, and operational control are different kinds of evidence. The geographical location of a sending device does not establish the identity or location of the person directing it.
On 21 October 2016, attacks against the managed DNS provider Dyn interfered with access to numerous well-known online services. Mirai-associated traffic was involved. Many affected organizations were not themselves the direct target; their users experienced disruption because name resolution depended on a shared provider. The incident made dependency concentration a central part of public discussion about DDoS risk.
On 28 February 2018, GitHub experienced a memcached-amplification attack reported at approximately 1.35 terabits per second. In 2023, HTTP/2 Rapid Reset incidents drew attention to application request rates rather than bandwidth alone. Such milestones should not be treated as a single, directly comparable sequence of records. Reports may describe an attempted flood, traffic observed at a mitigation edge, or traffic delivered to a victim, and may use different averaging intervals and aggregation methods. Historical significance can therefore lie in a new mechanism or exposed dependency even when a later attack has a larger headline figure.
Detection, measurement, and attribution
Detection begins by identifying which part of a service is deteriorating. Increased latency, packet loss, connection failures, rejected requests, and resource saturation provide different clues. A full Internet link suggests a different problem from an idle front end waiting on an overloaded database. Observations from only one layer can be misleading: a server may record few requests precisely because congestion is preventing requests from reaching it.
Common evidence includes packet samples, flow records, application logs, connection-state counts, and resource measurements. Flow records summarize communications using attributes such as endpoints, ports, protocol, duration, and byte counts, often without retaining full payloads. Application telemetry can reveal which functions are expensive and which responses are failing. External measurements help distinguish a local monitoring problem from a real loss of availability, while observations from multiple regions can expose partial disruption.
Three common attack-rate measures are bits per second, packets per second, and requests per second. Each describes a different burden. Bits per second concerns transmission volume; packets per second concerns per-packet handling; requests per second concerns application work, although request costs can vary dramatically. Concurrent connections, stream-creation rates, and completed transactions can be equally important. Reports should distinguish peak from sustained values and attempted traffic from successfully processed work.
Distinguishing malicious demand from legitimate popularity is difficult because both can produce valid requests from many sources. Indicators may include unusual timing, concentration on costly endpoints, high failure or cancellation rates, inconsistent client behavior, and an abrupt change in traffic composition. None is conclusive in isolation. Shared corporate proxies, mobile networks, accessibility software, and automated integrations can resemble abusive clients if assessed using simplistic rules.
Anomaly detection and machine learning can combine these signals, but their usefulness depends on representative data and operational context. Seasonal demand, product launches, and changes in client software can shift the baseline. Models can also fail when an attacker deliberately resembles normal traffic. Precision and recall matter operationally: excessive false positives deny service to legitimate users, while false negatives allow the attack to continue.
Attribution is a separate and usually harder task. An observed source address may belong to a compromised device, a reflector, a proxy, or rented infrastructure. Spoofing can make it an address from which no attack packet was actually sent. Stronger attribution may require correlations among command infrastructure, account records, financial transactions, seized systems, and other evidence unavailable in ordinary traffic logs. Technical descriptions should therefore distinguish the mechanism of an incident from claims about its operator or sponsor.
Filtering, traffic distribution, and mitigation architecture
DDoS mitigation works by reducing hostile demand, increasing effective capacity, or preventing expensive work from being allocated too early. The decisive question is where intervention occurs relative to the bottleneck. Filtering on a server cannot recover an upstream link already filled with attack traffic. In that case, packets must be discarded or diverted before they traverse the constrained connection, often with assistance from an Internet service provider or specialized mitigation network.
A content delivery network can absorb or answer some requests close to users, reducing demand on the origin service. Distributed reverse proxies can terminate connections and enforce controls at a larger number of locations. Anycast routing allows multiple sites to advertise the same service address, distributing incoming traffic according to network routing decisions. These mechanisms increase geographic and operational spread, but their effectiveness depends on topology, traffic distribution, and whether attackers can reach an unprotected origin directly.
Traffic scrubbing diverts traffic through systems that attempt to separate legitimate packets or requests from attack traffic. Diversion may be persistent or activated during an incident, using routing changes or proxy arrangements. Cleaned traffic is then delivered to the protected service. Border Gateway Protocol mechanisms can also support remotely triggered blackholing, which discards traffic toward a destination to protect surrounding infrastructure. Blackholing sacrifices reachability to that destination; it contains damage rather than preserving the targeted service.
At the application layer, rate limiting bounds how quickly work may be requested. Token-bucket and related mechanisms permit controlled bursts while limiting longer-term consumption. Limits can apply to accounts, network origins, operations, or combinations of signals. Per-address limits alone are imperfect because many legitimate users may share an address, while a distributed attacker may use many addresses. Cost-aware limits seek to reflect the actual processing burden rather than merely counting requests.
Challenges and web-application filtering can screen some automated traffic, but they introduce trade-offs. A challenge may obstruct users with disabilities, privacy tools, restricted browsers, or non-browser API clients. A web application firewall can recognize known patterns and impose request controls, yet it cannot reliably infer intent from every valid application transaction. Filtering rules are therefore often adjusted as the traffic and service conditions change.
Protocol and software changes remove specific asymmetries: SYN cookies reduce initial TCP state allocation, handshake cookies can defer expensive work until reachability is demonstrated, and patches can eliminate crash conditions or unsafe stream handling. No single measure covers every attack class. Effective mitigation combines controls across transport, application, and upstream networks while observing whether legitimate transactions continue to succeed, rather than equating a falling traffic counter with restored service.
Resilient design, economic effects, and governance
Resilience concerns how a system behaves when some hostile traffic gets through. Redundancy is useful only when components do not share the same vulnerable dependency or resource limit. Two application instances behind one saturated connection are not independent protection. Likewise, multiple regions can remain jointly dependent on a single DNS provider, identity service, control plane, or account-management system. The structure of dependencies matters as much as the number of servers.
Application architecture can contain overload through bounded queues, workload isolation, and load shedding. Instead of accepting unlimited work and failing unpredictably, a service can reject some requests promptly or reduce optional functionality. Circuit breakers limit repeated calls to an unhealthy dependency. Separating costly operations from essential ones can preserve basic service when a report generator or search function is overwhelmed. These measures address both malicious and accidental overload without assuming that every request can be classified correctly.
Retries are an important source of secondary load. When clients immediately repeat failed requests, a partially overloaded service can receive more work precisely when its capacity has fallen. Similar effects occur when one failing dependency causes multiple upstream services to retry. Retry limits, randomized delays, and explicit overload signals can reduce such amplification. However, an attacker need not obey these signals, so cooperative overload handling complements rather than replaces adversarial traffic controls.
Cloud computing changes the economics but does not abolish resource limits. Automatic scaling may increase capacity while also increasing expenditure, and scaling can be slower than the onset of an attack. Provider quotas, regional limits, and bottlenecks in databases or external APIs can prevent extra application instances from restoring availability. Economic denial of sustainability describes situations in which abusive demand imposes substantial operating costs even when the service remains technically accessible.
Preparedness and incident analysis require authorized testing, clear escalation paths, and evidence preservation. Load tests measure behavior under specified conditions; they do not establish immunity to all denial-of-service mechanisms. Tests can also affect shared networks or external dependencies, making the scope of authorization important. After an incident, separating attack-induced failures from defensive side effects helps identify whether outages resulted from resource exhaustion, overbroad filtering, routing changes, or interacting causes.
Unauthorized disruption is addressed by computer-misuse laws in many jurisdictions, including provisions concerning intentional impairment of computer systems. Cross-border infrastructure and compromised intermediaries complicate investigation and enforcement. At the standards and operational levels, continuing issues include incentives for source-address validation, secure defaults in connected devices, and comparable attack measurements. Protocol development also increasingly examines when a system allocates state or performs expensive computation, since preventing disproportionate work at the design stage can remove entire classes of denial-of-service opportunity.