Protecting Internet Key Exchange Protocol version 2 (IKEv2) Implementations from Distributed Denial of Service Attacks

Denial of Service (DoS) attacks have always been considered a serious threat. These attacks are usually difficult to defend against since the amount of resources the victim has is always bounded (regardless of how high it is) and because some resources are required for distinguishing a legitimate session from an attack. The Internet Key Exchange protocol version 2 (IKEv2) described in includes defense against DoS attacks. In particular, there is a cookie mechanism that allows the IKE Responder to defend itself against DoS attacks from spoofed IP-addresses. However, bot-nets have become widespread, allowing attackers to perform Distributed Denial of Service (DDoS) attacks, which are more difficult to defend against. This document presents recommendations to help the Responder counter (D)DoS attacks. It also introduces a new mechanism -- "puzzles" -- that can help accomplish this task.

The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in .

The IKE_SA_INIT Exchange described in Section 1.2 of involves the Initiator sending a single message. The Responder replies with a single message and also allocates memory for a structure called a half-open IKE Security Association (SA). This half-open SA is later authenticated in the IKE_AUTH Exchange. If that IKE_AUTH request never comes, the half-open SA is kept for an unspecified amount of time. Depending on the algorithms used and implementation, such a half-open SA will use from around 100 bytes to several thousands bytes of memory. This creates an easy attack vector against an IKE Responder. Generating the IKE_SA_INIT request is cheap. Sending large amounts of IKE_SA_INIT requests can cause a Responder to use up all its resources. If the Responder tries to defend against this by throttling new requests, this will also prevent legitimate Initiators from setting up IKE SAs. An obvious defense, which is described in , is limiting the number of half-open SAs opened by a single peer. However, since all that is required is a single packet, an attacker can use multiple spoofed source IP addresses. If we break down what a Responder has to do during an initial exchange, there are three stages: When the IKE_SA_INIT request arrives, the Responder: Generates or re-uses a Diffie-Hellman (D-H) private part. Generates a Responder Security Parameter Index (SPI). Stores the private part and peer public part in a half-open SA database. When the IKE_AUTH request arrives, the Responder: Derives the keys from the half-open SA. Decrypts the request. If the IKE_AUTH request decrypts properly: Validates the certificate chain (if present) in the IKE_AUTH request. The fourth stage where the Responder creates the Child SA is not reached by attackers who cannot pass the authentication step. Stage #1 is pretty light on CPU power, but requires some storage, and it's very light for the Initiator as well. Stage #2 includes private-key operations, so it is much heavier CPU-wise. Stage #3 may include public key operations if certificates are involved. These operations are often more computationly expensive than those performed at stage #2. To attack such a Responder, an attacker can attempt either to exhaust memory or to exhaust CPU. Without any protection, the most efficient attack is to send multiple IKE_SA_INIT requests and exhaust memory. This is easy because IKE_SA_INIT requests are cheap. There are obvious ways for the Responder to protect itself without changes to the protocol. It can reduce the time that an entry remains in the half-open SA database, and it can limit the amount of concurrent half-open SAs from a particular address or prefix. The attacker can overcome this by using spoofed source addresses. The stateless cookie mechanism from Section 2.6 of prevents an attack with spoofed source addresses. This doesn't completely solve the issue, but it makes the limiting of half-open SAs by address or prefix work. Puzzles, introduced in , accomplish the same thing only more of it. They make it harder for an attacker to reach the goal of getting a half-open SA. Puzzles do not have to be so hard that an attacker cannot afford to solve a single puzzle; it is enough that puzzles increase the cost of creating a half-open SAs, so the attacker is limited in the amount they can create. Reducing the lifetime of an abandoned half-open SA also reduces the impact of such attacks. For example, if a half-open SA is kept for 1 minute and the capacity is 60 thousand half-open SAs, an attacker would need to create one thousand half-open SAs per second. If the retention time is reduced to 3 seconds, the attacker would need to create 20 thousand half-open SAs per second to get the same result. By introducing a puzzle, each half-open SA becomes more expensive for an attacker, making it more likely to prevent an exhaustion attack against Responder memory. At this point, filling up the half-open SA database is no longer the most efficient DoS attack. The attacker has two alternative attacks to do better: Go back to spoofed addresses and try to overwhelm the CPU that deals with generating cookies, or Take the attack to the next level by also sending an IKE_AUTH request. If an attacker is so powerfull that it is able to overwhelm the Responder's CPU that deals with generating cookies, then the attack cannot be dealt with at the IKE level and must be handled by means of the Intrusion Prevention System (IPS) technology. On the other hand, the second alternative of sending an IKE_AUTH request is very cheap. It requires generating a proper IKE header with the correct IKE SPIs and a single Encrypted payload. The content of the payload is irrelevant and might be junk. The Responder has to perform the relatively expensive key derivation, only to find that the MAC on the Encrypted payload on the IKE_AUTH request fails the integrity check. If a Responder does not hold on to the calculated SKEYSEED and SK_* keys (which it should in case a valid IKE_AUTH comes in later) this attack might be repeated on the same half-open SA. Puzzles make attacks of such sort more costly for an attacker. See for details. Here too, the number of half-open SAs that the attacker can achieve is crucial, because each one allows the attacker to waste some CPU time. So making it hard to make many half-open SAs is important. A strategy against DDoS has to rely on at least 4 components: Hardening the half-open SA database by reducing retention time. Hardening the half-open SA database by rate-limiting single IPs/prefixes. Guidance on what to do when an IKE_AUTH request fails to decrypt. Increasing the cost of half-open SAs up to what is tolerable for legitimate clients. Puzzles are used as a solution for strategy #4.

As a UDP-based protocol, IKEv2 has to deal with packet loss through retransmissions. Section 2.4 of recommends "that messages be retransmitted at least a dozen times over a period of at least several minutes before giving up". Many retransmission policies in practice wait one or two seconds before retransmitting for the first time. Because of this, setting the timeout on a half-open SA too low will cause it to expire whenever even one IKE_AUTH request packet is lost. When not under attack, the half-open SA timeout SHOULD be set high enough that the Initiator will have enough time to send multiple retransmissions, minimizing the chance of transient network congestion causing an IKE failure. When the system is under attack, as measured by the amount of half-open SAs, it makes sense to reduce this lifetime. The Responder should still allow enough time for the round-trip, enough time for the Initiator to derive the D-H shared value, and enough time to derive the IKE SA keys and the create the IKE_AUTH request. Two seconds is probably as low a value as can realistically be used. It could make sense to assign a shorter value to half-open SAs originating from IP addresses or prefixes that are considered suspect because of multiple concurrent half-open SAs.

Even with DDoS, the attacker has only a limited amount of nodes participating in the attack. By limiting the amount of half-open SAs that are allowed to exist concurrently with each such node, the total amount of half-open SAs is capped, as is the total amount of key derivations that the Responder is forced to complete. In IPv4 it makes sense to limit the number of half-open SAs based on IP address. Most IPv4 nodes are either directly attached to the Internet using a routable address or are hidden behind a NAT device with a single IPv4 external address. For IPv6, ISPs assign between a /48 and a /64, so it does not make sense for rate-limiting to work on single IPv6 IPs. Instead, ratelimits should be done based on either the /48 or /64 of the misbehaving IPv6 address observed. The number of half-open SAs is easy to measure, but it is also worthwhile to measure the number of failed IKE_AUTH exchanges. If possible, both factors should be taken into account when deciding which IP address or prefix is considered suspicious. There are two ways to rate-limit a peer address or prefix: Hard Limit - where the number of half-open SAs is capped, and any further IKE_SA_INIT requests are rejected. Soft Limit - where if a set number of half-open SAs exist for a particular address or prefix, any IKE_SA_INIT request will be required to solve a puzzle. The advantage of the hard limit method is that it provides a hard cap on the amount of half-open SAs that the attacker is able to create. The disadvantage is that it allows the attacker to block IKE initiation from small parts of the Internet. For example, if an network service provider or some establishment offers Internet connectivity to its customers or employees through an IPv4 NAT device, a single malicious customer can create enough half-open SAs to fill the quota for the NAT device external IP address. Legitimate Initiators on the same network will not be able to initiate IKE. The advantage of a soft limit is that legitimate clients can always connect. The disadvantage is that an adversary with sufficient CPU resources can still effectively DoS the Responder. Regardless of the type of rate-limiting used, legitimate initiators that are not on the same network segments as the attackers will not be affected. This is very important as it reduces the adverse impact caused by the measures used to counteract the attack, and allows most initiators to keep working even if they do not support puzzles.

Section 2.6 of offers a mechanism to mitigate DoS attacks: the stateless cookie. When the server is under load, the Responder responds to the IKE_SA_INIT request with a calculated "stateless cookie" - a value that can be re-calculated based on values in the IKE_SA_INIT request without storing Responder-side state. The Initiator is expected to repeat the IKE_SA_INIT request, this time including the stateless cookie. This mechanism prevents DoS attacks from spoofed IP addresses, since an attacker needs to have a routable IP address to return the cookie. Attackers that have multiple source IP addresses with return routability, such as in the case of bot-nets, can fill up a half-open SA table anyway. The cookie mechanism limits the amount of allocated state to the number of attackers, multiplied by the number of half-open SAs allowed per peer address, multiplied by the amount of state allocated for each half-open SA. With typical values this can easily reach hundreds of megabytes.

The puzzle introduced here extends the cookie mechanism of . It is loosely based on the proof-of-work technique used in Bitcoins . Puzzles set an upper bound, determined by the attacker's CPU, to the number of negotiations the attacker can initiate in a unit of time. A puzzle is sent to the Initiator in two cases: The Responder is so overloaded that no half-open SAs may be created without solving a puzzle, or The Responder is not too loaded, but the rate-limiting method described in prevents half-open SAs from being created with this particular peer address or prefix without first solving a puzzle. When the Responder decides to send the challenge to solve a puzzle in response to a IKE_SA_INIT request, the message includes at least three components: Cookie - this is calculated the same as in , i.e. the process of generating the cookie is not specified. Algorithm, this is the identifier of a Pseudo-Random Function (PRF) algorithm, one of those proposed by the Initiator in the SA payload. Zero Bit Count (ZBC). This is a number between 8 and 255 (or a special value - 0, see ) that represents the length of the zero-bit run at the end of the output of the PRF function calculated over the cookie that the Initiator is to send. The values 1-8 are explicitly excluded, because they create a puzzle that is too easy to solve. Since the mechanism is supposed to be stateless for the Responder, either the same ZBC is used for all Initiators, or the ZBC is somehow encoded in the cookie. If it is global then it means that this value is the same for all the Initiators who are receiving puzzles at any given point of time. The Responder, however, may change this value over time depending on its load. Upon receiving this challenge, the Initiator attempts to calculate the PRF output using different keys. When enough keys are found such that the resulting PRF output calculated using each of them has a sufficient number of trailing zero bits, that result is sent to the Responder. The reason for using several keys in the results, rather than just one key, is to reduce the variance in the time it takes the initiator to solve the puzzle. We have chosen the number of keys to be four (4) as a compromise between the conflicting goals of reducing variance and reducing the work the Responder needs to perform to verify the puzzle solution. When receiving a request with a solved puzzle, the Responder verifies two things: That the cookie is indeed valid. That the results of PRF of the transmitted cookie calculated with the transmitted keys has a sufficient number of trailing zero bits. Example 1: Suppose the calculated cookie is 739ae7492d8a810cf5e8dc0f9626c9dda773c5a3 (20 octets), the algorithm is PRF-HMAC-SHA256, and the required number of zero bits is 18. After successively trying a bunch of keys, the Initiator finds the following four 3-octet keys that work: Key Last 32 Hex PRF Digits # 0-bits 061840e4f957b859d7fb1343b7b94a816c000018 0733240d4233d6278c96e3369227a07580000023 0c8a2a952a35d39d5ba06709da43af4070000020 0d94c85a0452b21571e401a3d00803679c000018 Example 2: Same cookie, but modify the required number of zero bits to 22. The first 4-octet keys that work to satisfy that requirement are 005d9e57, 010d8959, 0110778d, and 01187e37. Finding these requires 18,382,392 invocations of the PRF. # 0-bits Time to Find 4 keys (seconds) 8 0.0025 10 0.0078 12 0.0530 14 0.2521 16 0.8504 17 1.5938 18 3.3842 19 3.8592 2010.8876 The figures above were obtained on a 2.4 GHz single core i5. Run times can be halved or quartered with multi-core code, but would be longer on mobile phone processors, even if those are multi-core as well. With these figures 18 bits is believed to be a reasonable choice for puzzle level difficulty for all Initiators, and 20 bits is acceptable for specific hosts/prefixes. Using puzzles mechanism in the IKE_SA_INIT exchange is described in .

When the Responder is under attack, it SHOULD prefer previously authenticated peers who present a Session Resumption ticket . However, the Responder SHOULD NOT serve resumed Initiators exclusively because dropping all IKE_SA_INIT requests would lock out legitimate Initiators that have no resumption ticket. When under attack the Responder SHOULD require Initiators presenting Session Resumption Tickets to pass a return routability check by including the COOKIE notification in the IKE_SESSION_RESUME response message, as described in Section 4.3.2. of . Note that the Responder SHOULD cache tickets for a short time to reject reused tickets (Section 4.3.1), and therefore there should be no issue of half-open SAs resulting from replayed IKE_SESSION_RESUME messages. Several kinds of DoS attacks are possible on servers supported IKE Session Resumption. See Section 9.3 of for details.

Once the IKE_SA_INIT exchange is finished, the Responder is waiting for the first message of the IKE_AUTH exchange from the Initiator. At this point the Initiator is not yet authenticated, and this fact allows an attacker to perform an attack, described in . The attacker can just send garbage in the IKE_AUTH message forcing the Responder to perform costly CPU operations to compute SK_* keys. If the received IKE_AUTH message failed to decrypt correctly (or failed to pass ICV check), then the Responder SHOULD still keep the computed SK_* keys, so that if it happened to be an attack, then an attacker cannot get advantage of repeating the attack multiple times on a single IKE SA. The responder can also use puzzles in the IKE_AUTH exchange as decribed in .

In IKEv2 each side may use the "Hash and URL" Certificate Encoding to instruct the peer to retrieve certificates from the specified location (see Section 3.6 of for details). Malicious initiators can use this feature to mount a DoS attack on the responder by providing an URL pointing to a large file possibly containing garbage. While downloading the file the responder consumes CPU, memory and network bandwidth. To prevent this kind of attack, the responder should not blindly download the whole file. Instead, it SHOULD first read the initial few bytes, decode the length of the ASN.1 structure from these bytes, and then download no more than the decoded number of bytes. Note, that it is always possible to determine the length of ASN.1 structures used in IKEv2, if they are DER-encoded, by analyzing the first few bytes. However, since the content of the file being downloaded can be under the attacker's control, implementations should not blindly trust the decoded length and SHOULD check whether it makes sense before continuing to download the file. Implementations SHOULD also apply a configurable hard limit to the number of pulled bytes and SHOULD provide an ability for an administrator to either completely disable this feature or to limit its use to a configurable list of trusted URLs.

IKE Fragmentation described in allows IKE peers to avoid IP fragmentation of large IKE messages. Attackers can mount several kinds of DoS attacks using IKE Fragmentation. See Section 5 of for details on how to mitigate these attacks.

Once an IKE SA is created there usually are only a limited amount of IKE messages exchanged. This IKE traffic consists of exchanges aimed to create additional Child SAs, IKE rekeys, IKE deletions and IKE liveness tests. Some of these exchanges require relatively little resources (like liveness check), while others may be resource consuming (like creating or rekeying Child SA with D-H exchange). Since any endpoint can initiate a new exchange, there is a possibility that a peer would initiate too many exchanges that could exhaust host resources. For example, the peer can perform endless continuous Child SA rekeying or create an overwhelming number of Child SAs with the same Traffic Selectors etc. Such behavior can be caused by broken implementations, misconfiguration, or as an intentional attack. The latter becomes more of a real threat if the peer uses NULL Authentication, as described in . In this case the peer remains anonymous, allowing it to escape any responsibility for its behaviour. See Section 3 of for details on how to mitigate attacks when using NULL Authentication. The following recommendations apply especially for NULL Authenticated IKE sessions, but also apply to authenticated IKE sessions, with the difference that in the latter case, the identified peer can be locked out. If the IKEv2 window size is greater than one, peers are able to initiate multiple simultaneous exchanges that increase host resource consumption. Since there is no way in IKEv2 to decrease window size once it has been increased (see Section 2.3 of ), the window size cannot be dynamically adjusted depending on the load. It is NOT RECOMMENDED to allow an IKEv2 window size greater than one when NULL Authentication has been used. If a peer initiates an abusive amount of CREATE_CHILD_SA exchanges to rekey IKE SAs or Child SAs, the Responder SHOULD reply with TEMPORARY_FAILURE notifications indicating the peer must slow down their requests. If a peer creates many Child SA with the same or overlapping Traffic Selectors, implementations MAY respond with the NO_ADDITIONAL_SAS notification. If a peer initiates many exchanges of any kind, the Responder MAY introduce an artificial delay before responding to each request message. This delay would decrease the rate the Responder needs to process requests from any particular peer, and frees up resources on the Responder that can be used for answering legitimate clients. If the Responder receives retransmissions of the request message during the delay period, the retransmitted messages MUST be silently discarded. The delay must be short enough to avoid legitimate peers deleting the IKE SA due to a timeout. It is believed that a few seconds is enough. Note however, that even a few seconds may be too long when settings rely on an immediate response to the request message, e.g. for the purposes of quick detection of a dead peer. If these counter-measures are inefficient, implementations MAY delete the IKE SA with an offending peer by sending Delete Payload. In IKE, a client can request various configuration attributes from server. Most often these attributes include internal IP addresses. Malicious clients can try to exhaust a server's IP address pool by continuously requesting a large number of internal addresses. Server implementations SHOULD limit the number of IP addresses allocated to any particular client. Note, this is not possible with clients using NULL Authentication, since their identity cannot be verified.

This section outlines a plan for defending a Responder from a DDoS attack based on the techniques described earlier. The numbers given here are not normative, and their purpose is to illustrate the configurable parameters needed for surviving DDoS attacks. Implementations are deployed in different environments, so it is RECOMMENDED that the parameters be settable. For example, most commercial products are required to undergo benchmarking where the IKE SA establishment rate is measured. Benchmarking is indistinguishable from a DoS attack and the defenses described in this document may defeat the benchmark by causing exchanges to fail or take a long time to complete. Parameters SHOULD be tunable to allow for benchmarking (if only by turning DDoS protection off). Since all countermeasures may cause delays and additional work for the Initiators, they SHOULD NOT be deployed unless an attack is likely to be in progress. To minimize the burden imposed on Initiators, the Responder should monitor incoming IKE requests, for two scenarios: A general DDoS attack. Such an attack is indicated by a high number of concurrent half-open SAs, a high rate of failed IKE_AUTH exchanges, or a combination of both. For example, consider a Responder that has 10,000 distinct peers of which at peak 7,500 concurrently have VPN tunnels. At the start of peak time, 600 peers might establish tunnels within any given minute, and tunnel establishment (both IKE_SA_INIT and IKE_AUTH) takes anywhere from 0.5 to 2 seconds. For this Responder, we expect there to be less than 20 concurrent half-open SAs, so having 100 concurrent half-open SAs can be interpreted as an indication of an attack. Similarly, IKE_AUTH request decryption failures should never happen. Supposing that the tunnels are established using EAP (see Section 2.16 of ), users may be expected to enter a wrong password about 20% of the time. So we'd expect 125 wrong password failures a minute. If we get IKE_AUTH decryption failures from multiple sources more than once per second, or EAP failures more than 300 times per minute, this can also be an indication of a DDoS attack. An attack from a particular IP address or prefix. Such an attack is indicated by an inordinate amount of half-open SAs from a specific IP address or prefix, or an inordinate amount of IKE_AUTH failures. A DDoS attack may be viewed as multiple such attacks. If these are mitigated successfully, there will not be a need to enact countermeasures on all Initiators. For example, measures might be 5 concurrent half-open SAs, 1 decrypt failure, or 10 EAP failures within a minute. Note that using counter-measures against an attack from a particular IP address may be enough to avoid the overload on the half-open SA database. In this case the number of failed IKE_AUTH exchanges will never exceed the threshold of attack detection. When there is no general DDoS attack, it is suggested that no cookie or puzzles be used. At this point the only defensive measure is to monitor the number of half-open SAs, and setting a soft limit per peer IP or prefix. The soft limit can be set to 3-5. If the puzzles are used, the puzzle difficulty should be set to such a level (number of zero-bits) that all legitimate clients can handle it without degraded user experience. As soon as any kind of attack is detected, either a lot of initiations from multiple sources or a lot of initiations from a few sources, it is best to begin by requiring stateless cookies from all Initiators. This will This will mitigate attacks based on IP address spoofing, and help avoid the need to impose a greater burden in the form of puzzles on the general population of Initiators. This makes the per-node or per-prefix soft limit more effective. When cookies are activated for all requests and the attacker is still managing to consume too many resources, the Responder MAY start to use puzzles for these requests or increase the difficulty of puzzles imposed on IKE_SA_INIT requests coming from suspicious nodes/prefixes. This should still be doable by all legitimate peers, but the use of puzzles at a higher difficulty may degrade the user experience, for example by taking up to 10 seconds to solve the puzzle. If the load on the Responder is still too great, and there are many nodes causing multiple half-open SAs or IKE_AUTH failures, the Responder MAY impose hard limits on those nodes. If it turns out that the attack is very widespread and the hard caps are not solving the issue, a puzzle MAY be imposed on all Initiators. Note that this is the last step, and the Responder should avoid this if possible.

This section describes how the puzzle mechanism is used in IKEv2. It is organized as follows. The describes using puzzles in the IKE_SA_INIT exchange and the describes using puzzles in the IKE_AUTH exchange. Both sections are divided into subsections describing how puzzles should be presented, solved and processed by the Initiator and the Responder.

IKE Initiator indicates the desire to create a new IKE SA by sending an IKE_SA_INIT request message. The message may optionally contain a COOKIE notification if this is a repeated request performed after the Responder's demand to return a cookie.

]]> According to the plan, described in , the IKE Responder should monitor incoming requests to detect whether it is under attack. If the Responder learns that a (D)DoS attack is likely to be in progress, then its actions depend on the volume of the attack. If the volume is moderate, then the Responder requests the Initiator to return a cookie. If the volume is so high, that puzzles need to be used for defense, then the Responder requests the Initiator to solve a puzzle. The Responder MAY choose to process some fraction of IKE_SA_INIT requests without presenting a puzzle while being under attack to allow legacy clients, that don't support puzzles, to have a chance to be served. The decision whether to process any particular request must be probabilistic, with the probability depending on the Responder's load (i.e. on the volume of attack). The requests that don't contain the COOKIE notification MUST NOT participate in this lottery. In other words, the Responder must first perform a return routability check before allowing any legacy client to be served if it is under attack. See for details.

If the Responder makes a decision to use puzzles, then it includes two notifications in its response message - the COOKIE notification and the PUZZLE notification. Note that the PUZZLE notification MUST always be accompanied with the COOKIE notification, since the content of the COOKIE notification is used as an input data when solving puzzle. The format of the PUZZLE notification is described in .

The presence of these notifications in an IKE_SA_INIT response message indicates to the Initiator that it should solve the puzzle to have a better chance to be served.

The PUZZLE notification contains the difficulty level of the puzzle - the minimum number of trailing zero bits that the result of PRF must contain. In diverse environments it is next to impossible for the Responder to set any specific difficulty level that will result in roughly the same amount of work for all Initiators, because computation power of different Initiators may vary by an order of magnitude, or even more. The Responder may set the difficulty level to 0, meaning that the Initiator is requested to spend as much power to solve a puzzle as it can afford. In this case no specific value of ZBC is required from the Initiator, however the larger the ZBC that Initiator is able to get, the better the chance is that it will be served by the Responder. In diverse environments it is RECOMMENDED that the Initiator set the difficulty level to 0, unless the attack volume is very high. If the Responder sets a non-zero difficulty level, then the level should be determined by analyzing the volume of the attack. The Responder MAY set different difficulty levels to different requests depending on the IP address the request has come from.

The PUZZLE notification also contains identifier of the algorithm, that must be used by Initiator to compute puzzle. Cryptographic algorithm agility is considered an important feature for modern protocols (). Algorithm agility ensures that a protocol doesn't rely on a single built-in set of cryptographic algorithms, but has a means to replace one set with another, and negotiate new algorithms with the peer. IKEv2 fully supports cryptographic algorithm agility for its core operations. To support crypto agility in case of puzzles, the algorithm that is used to compute a puzzle needs to be negotiated during the IKE_SA_INIT exchange. The negotiation is performed as follows. The initial request message sent by the Initiator contains an SA payload with the list of transforms the Initiator supports and is willing to use in the IKE SA being established. The Responder parses the received SA payload and finds a mutually supported PRFs. The Responder selects the preferred PRF from the list of mutually supported ones and includes it into the PUZZLE notification. There is no requirement that the PRF selected for puzzles be the same as the PRF that is negotiated later for use in core IKE SA crypto operations. If there are no mutually supported PRFs, then IKE SA negotiation will fail anyway and there is no reason to return a puzzle. In this case the Responder returns a NO_PROPOSAL_CHOSEN notification. Note that PRF is a mandatory transform type for IKE SA (see Sections 3.3.2 and 3.3.3 of ) and at least one transform of this type must always be present in the SA payload in an IKE_SA_INIT request message.

If the Responder supports puzzles then a cookie should be computed in such a manner that the Responder is able to learn some important information from the sole cookie, when it is later returned back by Initiator. In particular - the Responder should be able to learn the following information: Whether the puzzle was given to the Initiator or only the cookie was requested. The difficulty level of the puzzle given to the Initiator. The number of consecutive puzzles given to the Initiator. The amount of time the Initiator spent to solve the puzzles. This can be calculated if the cookie is timestamped. This information helps the Responder to make a decision whether to serve this request or demand more work from the Initiator. One possible approach to get this information is to encode it in the cookie. The format of such encoding is an implementation detail of Responder, as the cookie would remain an opaque blob to the Initiator. If this information is encoded in the cookie, then the Responder MUST make it integrity protected, so that any intended or accidental alteration of this information in the returned cookie is detectable. So, the cookie would be generated as:

| | Hash(Ni | IPi | SPIi | | ) ]]> Note, that according to the Section 2.6 of , the size of the cookie cannot exceed 64 bytes. Alternatively, the Responder may continue to generate a cookie as suggested in Section 2.6 of , but associate the additional information, using local storage identified with the particular version of the secret. In this case the Responder should have different secrets for every combination of difficulty level and number of consecutive puzzles, and should change the secrets periodically, keeping a few previous versions, to be able to calculate how long ago a cookie was generated. The Responder may also combine these approaches. This document doesn't mandate how the Responder learns this information from a cookie.

If the Initiator receives a puzzle but it doesn't support puzzles, then it will ignore the PUZZLE notification as an unrecognized status notification (in accordance to Section 3.10.1 of ). The Initiator MAY ignore the PUZZLE notification if it is not willing to spend resources to solve the puzzle of the requested difficulty, even if it supports puzzles. In both cases the Initiator acts as described in Section 2.6 of - it restarts the request and includes the received COOKIE notification into it. The Responder should be able to distinguish the situation when it just requested a cookie from the situation where the puzzle was given to the Initiator, but the Initiator for some reason ignored it. If the received message contains a PUZZLE notification and doesn't contain a COOKIE notification, then this message is malformed because it requests to solve the puzzle, but doesn't provide enough information to allow the puzzle to be solved. In this case the Initiator MUST ignore the received message and continue to wait until either a valid PUZZLE notification is received or the retransmission timer fires. If it fails to receive a valid message after several retransmissions of IKE_SA_INIT requests, then it means that something is wrong and the IKE SA cannot be established. If the Initiator supports puzzles and is ready to solve them, then it tries to solve the given puzzle. After the puzzle is solved the Initiator restarts the request and returns back to the Responder the puzzle solution in a new payload called a Puzzle Solution payload (denoted as PS, see ) along with the received COOKIE notification.

]]>

General principals of constructing puzzles in IKEv2 are described in . They can be summarized as follows: given unpredictable string S and pseudo-random function PRF find N different keys Ki (where i=[1..N]) for that PRF so that the result of PRF(Ki,S) has at least the specified number of trailing zero bits. This specification requires that the solution to the puzzle contain 4 different keys (i.e. N=4). In the IKE_SA_INIT exchange it is the cookie that plays the role of unpredictable string S. In other words, in the IKE_SA_INIT the task for the IKE Initiator is to find the four different, equal-sized keys Ki for the agreed upon PRF such that each result of PRF(Ki,cookie) where i = [1..4] has a sufficient number of trailing zero bits. Only the content of the COOKIE notification is used in puzzle calculation, i.e. the header of the Notification payload is not included. Note, that puzzles in the IKE_AUTH exchange are computed differently than in the IKE_SA_INIT_EXCHANGE. See for details.

The received request must at least contain a COOKIE notification. Otherwise it is an initial request and it must be processed according to . First, the cookie MUST be checked for validity. If the cookie is invalid, then the request is treated as initial and is processed according to . It is RECOMMENDED that a new cookie is requested in this case. If the cookie is valid, then some important information is learned from it, or from local state based on identifier of the cookie's secret (see for details). This information helps the Responder to sort out incoming requests, giving more priority to those which were created by spending more of the Initiator's resources. First, the Responder determines if it requested only a cookie, or presented a puzzle to the Initiator. If no puzzle was given, this means that at the time the Responder requested a cookie it didn't detect the (D)DoS attack or the attack volume was low. In this case the received request message must not contain the PS payload, and this payload MUST be ignored if the message contains a PS payload for any reason. Since no puzzle was given, the Responder marks the request with the lowest priority since the Initiator spent little resources creating it. If the Responder learns from the cookie that the puzzle was given to the Initiator, then it looks for the PS payload to determine whether its request to solve the puzzle was honored or not. If the incoming message doesn't contain a PS payload, this means that the Initiator either doesn't support puzzles or doesn't want to deal with them. In either case the request is marked with the lowest priority since the Initiator spent little resources creating it. If a PS payload is found in the message, then the Responder MUST verify the puzzle solution that it contains. The solution is interpreted as four different keys. The result of using each of them in the PRF (as described in ) must contain at least the requested number of trailing zero bits. The Responder MUST check all of the four returned keys. If any checked result contains fewer bits than were requested, this means that the Initiator spent less resources than expected by the Responder. This request is marked with the lowest priority. If the Initiator provided the solution to the puzzle satisfying the requested difficulty level, or if the Responder didn't indicate any particular difficulty level (by setting ZBC to zero) and the Initiator was free to select any difficulty level it can afford, then the priority of the request is calculated based on the following considerations: The Responder must take the smallest number of trailing zero bits among the checked results and count it as the number of zero bits the Initiator solved for. The higher number of zero bits the Initiator provides, the higher priority its request should receive. The more consecutive puzzles the Initiator solved, the higher priority it should receive. The more time the Initiator spent solving the puzzles, the higher priority it should receive. After the priority of the request is determined the final decision whether to serve it or not is made.

The Responder decides what to do with the request based on the request's priority and the Responder's current load. There are three possible actions: Accept request. Reject request. Demand more work from the Initiator by giving it a new puzzle. The Responder SHOULD accept an incoming request if its priority is high - this means that the Initiator spent quite a lot of resources. The Responder MAY also accept some low-priority requests where the Initiators don't support puzzles. The percentage of accepted legacy requests depends on the Responder's current load. If the Initiator solved the puzzle, but didn't spend much resources for it (the selected puzzle difficulty level appeared to be low and the Initiator solved it quickly), then the Responder SHOULD give it another puzzle. The more puzzles the Initiator solves the higher its chances are to be served. The details of how the Responder makes a decision for any particular request are implementation dependent. The Responder can collect all of the incoming requests for some short period of time, sort them out based on their priority, calculate the number of available memory slots for half-open IKE SAs and then serve that number of requests from the head of the sorted list. The remainder of requests can be either discarded or responded to with new puzzle requests. Alternatively, the Responder may decide whether to accept every incoming request with some kind of lottery, taking into account its priority and the available resources.

Once the IKE_SA_INIT exchange is completed, the Responder has created a state and is waiting for the first message of the IKE_AUTH exchange from the Initiator. At this point the Initiator has already passed the return routability check and has proved that it has performed some work to complete IKE_SA_INIT exchange. However, the Initiator is not yet authenticated and this allows a malicious Initiator to perform an attack, described in . Unlike a DoS attack in the IKE_SA_INIT exchange, which is targeted on the Responder's memory resources, the goal of this attack is to exhaust a Responder's CPU power. The attack is performed by sending the first IKE_AUTH message containing garbage. This costs nothing to the Initiator, but the Responder has to perform relatively costly operations when computing the D-H shared secret and deriving SK_* keys to be able to verify authenticity of the message. If the Responder doesn't keep the computed keys after an unsuccessful verification of the IKE_AUTH message, then the attack can be repeated several times on the same IKE SA. The Responder can use puzzles to make this attack more costly for the Initiator. The idea is that the Responder includes a puzzle in the IKE_SA_INIT response message and the Initiator includes a puzzle solution in the first IKE_AUTH request message outside the Encrypted payload, so that the Responder is able to verify puzzle solution before computing the D-H shared secret. The difficulty level of the puzzle should be selected so that the Initiator would spend substantially more time to solve the puzzle than the Responder to compute the shared secret. The Responder should constantly monitor the amount of the half-open IKE SA states that receive IKE_AUTH messages that cannot be decrypted due to integrity check failures. If the percentage of such states is high and it takes an essential fraction of Responder's computing power to calculate keys for them, then the Responder may assume that it is under attack and SHOULD use puzzles to make it harder for attackers.

The Responder requests the Initiator to solve a puzzle by including the PUZZLE notification in the IKE_SA_INIT response message. The Responder MUST NOT use puzzles in the IKE_AUTH exchange unless a puzzle has been previously presented and solved in the preceding IKE_SA_INIT exchange.

The difficulty level of the puzzle in the IKE_AUTH exchange should be chosen so that the Initiator would spend more time to solve the puzzle than the Responder to compute the D-H shared secret and the keys needed to decrypt and verify the IKE_AUTH request message. On the other hand, the difficulty level should not be too high, otherwise legitimate clients will experience an additional delay while establishing the IKE SA. Note, that since puzzles in the IKE_AUTH exchange are only allowed to be used if they were used in the preceding IKE_SA_INIT exchange, the Responder would be able to estimate the computational power of the Initiator and select the difficulty level accordingly. Unlike puzzles in the IKE_SA_INIT, the requested difficulty level for IKE_AUTH puzzles MUST NOT be zero. In other words, the Responder must always set a specific difficulty level and must not let the Initiator to choose it on its own.

The algorithm for the puzzle is selected as described in . There is no requirement that the algorithm for the puzzle in the IKE_SA INIT exchange be the same as the algorithm for the puzzle in IKE_AUTH exchange; however, it is expected that in most cases they will be the same.

If the IKE_SA_INIT response message contains the PUZZLE notification and the Initiator supports puzzles, it MUST solve the puzzle. Note, that puzzle construction in the IKE_AUTH exchange differs from the puzzle construction in the IKE_SA_INIT exchange and is described in . Once the puzzle is solved the Initiator sends the IKE_AUTH request message, containing the Puzzle Solution payload.

]]> The Puzzle Solution (PS) payload MUST be placed outside the Encrypted payload, so that the Responder is able to verify the puzzle before calculating the D-H shared secret and the SK_* keys. If IKE Fragmentation is used in IKE_AUTH exchange, then the PS payload MUST be present only in the first IKE Fragment message, in accordance with the Section 2.5.3 of . Note, that calculation of the puzzle in the IKE_AUTH exchange doesn't depend on the content of the IKE_AUTH message (see ). Thus the Initiator has to solve the puzzle only once and the solution is valid for both unfragmented and fragmented IKE messages.

A puzzle in the IKE_AUTH exchange is computed differently than in the IKE_SA_INIT exchange (see ). The general principle is the same; the difference is in the construction of the string S. Unlike the IKE_SA_INIT exchange, where S is the cookie, in the IKE_AUTH exchange S is a concatenation of Nr and SPIr. In other words, the task for IKE Initiator is to find the four different keys Ki for the agreed upon PRF such that each result of PRF(Ki,Nr | SPIr) where i=[1..4] has a sufficient number of trailing zero bits. Nr is a nonce used by the Responder in the IKE_SA_INIT exchange, stripped of any headers. SPIr is the IKE Responder's SPI from the IKE header of the SA being established.

If the Responder requested the Initiator to solve a puzzle in the IKE_AUTH exchange, then it MUST silently discard all the IKE_AUTH request messages without the Puzzle Solution payload. Once the message containing a solution to the puzzle is received, the Responder MUST verify the solution before performing computationlly intensive operations i.e. computing the D-H shared secret and the SK_* keys. The Responder MUST verify all four of the returned keys. The Responder MUST silently discard the received message if any checked verification result is not correct (contains insufficient number of trailing zero bits). If the Responder successfully verifies the puzzle and calculates the SK_* key, but the message authenticity check fails, then it SHOULD save the calculated keys in the IKE SA state while waiting for the retransmissions from the Initiator. In this case the Responder may skip verification of the puzzle solution and ignore the Puzzle Solution payload in the retransmitted messages. If the Initiator uses IKE Fragmentation, then it is possible, that due to packet loss and/or reordering the Responder could receive non-first IKE Fragment messages before receiving the first one containing the PS payload. In this case the Responder MAY choose to keep the received fragments until the first fragment containing the solution to the puzzle is received. In this case the Responder SHOULD NOT try to verify authenticity of the kept fragments until the first fragment with the PS payload is received and the solution to the puzzle is verified. After successful verification of the puzzle, the Responder can then calculate the SK_* key and verify authenticity of the collected fragments.

The PUZZLE notification is used by the IKE Responder to inform the Initiator about the need to solve the puzzle. It contains the difficulty level of the puzzle and the PRF the Initiator should use.

Protocol ID (1 octet) -- MUST be 0. SPI Size (1 octet) - MUST be 0, meaning no Security Parameter Index (SPI) is present. Notify Message Type (2 octets) -- MUST be <TBA by IANA>, the value assigned for the PUZZLE notification. PRF (2 octets) -- Transform ID of the PRF algorithm that must be used to solve the puzzle. Readers should refer to the section "Transform Type 2 - Pseudo-Random Function Transform IDs" in for the list of possible values. Difficulty (1 octet) -- Difficulty Level of the puzzle. Specifies the minimum number of trailing zero bits (ZBC), that each of the results of PRF must contain. Value 0 means that the Responder doesn't request any specific difficulty level and the Initiator is free to select an appropriate difficulty level on its own (see for details). This notification contains no data.

The solution to the puzzle is returned back to the Responder in a dedicated payload, called the Puzzle Solution payload and denoted as PS in this document.

Puzzle Solution Data (variable length) -- Contains the solution to the puzzle - four different keys for the selected PRF. This field MUST NOT be empty. All of the keys MUST have the same size, therefore the size of this field is always a mutiple of 4 bytes. If the selected PRF accepts only fixed-size keys, then the size of each key MUST be of that fixed size. If the agreed upon PRF accepts keys of any size, then then the size of each key MUST be between 1 octet and the preferred key length of the PRF (inclusive). It is expected that in most cases the keys will be 4 (or even less) octets in length, however it depends on puzzle difficulty and on the Initiator's strategy to find solutions, and thus the size is not mandated by this specification. The Responder determines the size of each key by dividing the size of the Puzzle Solution Data by 4 (the number of keys). Note that the size of Puzzle Solution Data is the size of Payload (as indicated in Payload Length field) minus 4 - the size of Payload Header. The payload type for the Puzzle Solution payload is <TBA by IANA>.

The puzzle difficulty level should be set by balancing the requirement to minimize the latency for legitimate Initiators with making things difficult for attackers. A good rule of thumb is for taking about 1 second to solve the puzzle. A typical Initiator or bot-net member in 2014 can perform slightly less than a million hashes per second per core, so setting the difficulty level to n=20 is a good compromise. It should be noted that mobile Initiators, especially phones are considerably weaker than that. Implementations should allow administrators to set the difficulty level, and/or be able to set the difficulty level dynamically in response to load. Initiators should set a maximum difficulty level beyond which they won't try to solve the puzzle and log or display a failure message to the administrator or user.

Care must be taken when selecting parameters for the puzzles, in particular the puzzle difficulty. If the puzzles are too easy for the majority of attacker, then the puzzle mechanism wouldn't be able to prevent (D)DoS attacks and would only impose an additional burden on legitimate Initiators. On the other hand, if the puzzles are too hard for the majority of Initiators, then many legitimate users would experience unacceptable delays in IKE SA setup (and unacceptable power consumption on mobile devices), that might cause them to cancel the connection attempt. In this case the resources of the Responder are preserved, however the DoS attack can be considered successful. Thus a sensible balance should be kept by the Responder while choosing the puzzle difficulty - to defend itself and to not over-defend itself. It is RECOMMENDED that the puzzle difficulty be chosen so, that the Responder's load remains close to the maximum it can tolerate. It is also RECOMMENDED to dynamically adjust the puzzle difficulty in accordance to the current Responder's load. Solving puzzles requires a lot of CPU power that increases power consumption. This additional power consumption can negatively affect battery-powered Initiators, e.g. mobile phones or some IoT devices. If puzzles are too hard, then the required additional power consumption may appear to be unacceptable for some Initiators. The Responder SHOULD take this possibility into consideration while choosing the puzzle difficulty, and while selecting which percentage of Initiators are allowed to reject solving puzzles. See for details. If the Initiator uses NULL Authentication then its identity is never verified. This condition may be used by attackers to perform a DoS attack after the IKE SA is established. Responders that allow unauthenticated Initiators to connect must be prepared to deal with various kinds of DoS attacks even after the IKE SA is created. See for details. To prevent amplification attacks implementations must strictly follow the retransmission rules described in Section 2.1 of .

This document defines a new payload in the "IKEv2 Payload Types" registry:

Puzzle Solution PS ]]> This document also defines a new Notify Message Type in the "IKEv2 Notify Message Types - Status Types" registry:

PUZZLE ]]>

The authors thank Tero Kivinen, Yaron Sheffer, and Scott Fluhrer for their contributions to the design of the protocol. In particular, Tero Kivinen suggested the kind of puzzle where the task is to find a solution with a requested number of zero trailing bits. Yaron Sheffer and Scott Fluhrer suggested a way to make puzzle difficulty less erratic by solving several weaker puzles. The authors also thank David Waltermire and Paul Wouters for their careful reviews of the document, Graham Bartlett for pointing out to the possibility of the "Hash & URL" related attack, and all others who commented the document.