Analysis of Potential Solutions for Revealing a Host Identifier (HOST

As reported in , several issues are encountered when an IP address is shared among several subscribers. These issues are encountered in various deployment contexts, e.g., Carrier-Grade NAT (CGN), application proxies, or Address plus Port (A+P) . Examples of such issues are: implicit identification (Section 13.2 of ), spam (Section 13.3 of ), blacklisting a misbehaving host (Section 13.1 of ), or redirecting users with infected machines to a dedicated portal (Section 5.1 of ). In particular, some servers use the source IPv4 address as an identifier to treat some incoming connections differently. Due to the deployment of CGNs (e.g., NAT44 , NAT64 ), that address will be shared. In particular, when a server receives packets from the same source address, because this address is shared, the server does not know which host is the sending host . The sole use of the IPv4 address is not sufficient to uniquely distinguish a host. As a mitigation, it is tempting to investigate ways that would disclose information to be used by the remote server as a means of uniquely disambiguating packets sent from hosts using the same IPv4 address. The risk of not mitigating these issues include: OPEX (Operational Expenditure) increase for IP connectivity service providers (costs induced by calls to a hotline), revenue loss for content providers (loss of users' audience), and customers' dissatisfaction (low quality of experience, service segregation, etc.). The purpose of this document is to analyze a set of alternative channels to convey a host identifier and to assess to what extent the alternatives solve the problem described in . The evaluation is intended to be comprehensive, regardless of the maturity or validity of any currently known or proposed solution. The alternatives analyzed in the document are listed below: Use the Identification field of the IP header (denoted as IP-ID, ). Define a new IP option (). Define a new TCP option (). Inject application headers (). Enable Proxy Protocol (). Assign port sets (). Activate HIP (Host Identity Protocol) (). Use a notification channel (). Use an out-of-band mechanism (). A synthesis is provided in , while the detailed analysis is elaborated in . discusses privacy issues common to all proposed solutions. It is out of scope of this document to elaborate on privacy issues specific to each solution. This document does not include any recommendations because the working group felt that it was too premature to include one.

Policies that rely on source IP addresses and that are enforced by some servers will be applied to all hosts sharing the same IP address. For example, blacklisting the IP address of a spammer host will result in all other hosts that share that address having their access to the requested service restricted. describes the issues in detail. Therefore, due to address sharing, servers need extra information beyond the source IP address to differentiate the sending host. We call this information the HOST_ID. The HOST_ID identifies a host under a shared IP address. Privacy-related considerations are discussed in . Within this document, a host can be any computer located behind a Home Gateway or directly connected to an address-sharing function located in the network provider's domain (typically this would be the Home Gateway itself). Because the HOST_ID is used by a remote server to sort out the packets by sending host, the HOST_ID must be unique to each host under the same shared IP address, where possible. In the case where only the Home Gateway is revealed to the operator side of the translation function, the HOST_ID need only be unique to the Home Gateway. The HOST_ID does not need to be globally unique. Of course, the combination of the (public) IP source address and the identifier (i.e., HOST_ID) ends up being unique. If the HOST_ID is conveyed at the IP level, all packets will have to bear the identifier. If it is conveyed at a higher connection-oriented level, the identifier is only needed once in the session establishment phase (for instance, a TCP three-way handshake), then all packets received in this session will be attributed to the HOST_ID designated during the session opening. Within this document, we assume the operator-side address-sharing function injects the HOST_ID. Another deployment option to avoid potential performance degradation is to let the host or Home Gateway inject its HOST_ID, but the address-sharing function will check its content (just like an IP anti-spoofing function). For some proposals, the HOST_ID is retrieved using an out-of-band mechanism or signaled in a dedicated notification channel. For A+P and its variants, port set announcements may be needed as discussed in . Security considerations are common to all analyzed solutions (see ). Privacy-related aspects are discussed in . The HOST_ID can be ambiguous for hosts with multiple interfaces or multiple addresses assigned to a single interface. HOST_IDs that are the same may be used to imply or infer the same end system, but HOST_IDs that are different should not be used to imply or infer whether the end systems are the same or different.

IP address sharing is motivated by a number of different factors. For years, many network operators have conserved public IPv4 addresses by making use of Customer Premises Equipment (CPE) that assigns a single public IPv4 address to all hosts within the customer's local area network and uses NAT to translate between locally unique private IPv4 addresses and the CPE's public address. With the exhaustion of IPv4 address space, address sharing between customers on a much larger scale is likely to become much more prevalent. While many individual users are unaware of and uninvolved in decisions about whether their unique IPv4 addresses get revealed when they send data via IP, some users realize privacy benefits associated with IP address sharing, and some may even take steps to ensure that NAT functionality sits between them and the public Internet. IP address sharing makes the actions of all users behind the NAT function unattributable to any single host, creating room for abuse but also providing some identity protection for non&nbhy;abusive users who wish to transmit data with reduced risk of being uniquely identified. The proposals considered in this document help differentiate between hosts that share a public IP address. The extent of that differentiation depends on what information is included in the HOST_ID. The volatility of the HOST_ID information is similar to that of the internal IP address: a distinct HOST_ID may be used by the address-sharing function when the host reboots or gets a new internal IP address. As with persistent IP addresses, persistent HOST_IDs facilitate user tracking over time. As a general matter, the HOST_ID proposals do not seek to make hosts any more identifiable than they would be if they were using a public, non-shared IP address. However, depending on the solution proposal, the addition of HOST_ID information may allow a device to be fingerprinted more easily than it otherwise would be. To prevent this, the following design considerations are to be taken into account: It is recommended that HOST_IDs be limited to providing local uniqueness rather than global uniqueness. The address-sharing function should not use permanent HOST_ID values. Should multiple solutions be combined (e.g., TCP option and Forwarded header) that include different pieces of information in the HOST_ID, fingerprinting may become even easier. To prevent this, an address-sharing function that is able to inject HOST_IDs in several layers should reveal the same subsets of information at each layer. For example, if one layer references the lower 16 bits of an IPv4 address, the other layer should reference these 16 bits too. A HOST_ID can be spoofed, as this is also the case for spoofing an IP address. Furthermore, users of network-based anonymity services (like Tor [TOR]) may be capable of stripping HOST_ID information before it reaches its destination. In order to control the information revealed to external parties, an address-sharing function should be able to strip, rewrite, and add HOST_ID fields. An address-sharing function may be configured to enforce different end-user preferences with regards to HOST_ID injection. For example, HOST_ID injection can be disabled for some users. This feature is policy based and deployment specific. HOST_ID specification document(s) should explain the privacy impact of the solutions they specify, including the extent of HOST_ID uniqueness and persistence, assumptions made about the lifetime of the HOST_ID, whether and how the HOST_ID can be obfuscated or recycled, whether location information can be exposed, and the impact of the use of the HOST_ID on device or implementation fingerprinting. provides further guidance. For more discussion about privacy, refer to .

The IPv4 ID (Identification field of IP header, i.e., IP-ID) can be used to insert information that uniquely distinguishes a host among those sharing the same IPv4 address. Use of the IP-ID as a channel to convey the HOST_ID is a theoretical construct (i.e., it is an undocumented proposal). An address-sharing function can rewrite the IP-ID field to insert a value that is unique to the host (16 bits are sufficient to uniquely disambiguate hosts sharing the same IP address). The address-sharing function injecting the HOST_ID must follow the rules defined in ; in particular, the same HOST_ID is not reassigned to another host sharing the same IP address during a given time interval. A variant of this approach relies upon the format of certain packets, such as TCP SYN, where the IP-ID can be modified to contain a 16-bit HOST_ID. Address-sharing devices using this solution would be required to indicate that they do so, possibly using a special DNS record.

This usage is not consistent with the fragment reassembly use of the Identification field or the updated handling rules for the Identification field . Complications may arise if the packet is fragmented before reaching the device that is injecting the HOST_ID. To appropriately handle those packet fragments, the address-sharing function will need to maintain a lot of state. Another complication to be encountered is where translation is balanced among several NATs; setting the appropriate HOST_ID by a given NAT would alter the coordination between those NATs. Of course, one can argue that this coordinated NAT scenario is not a typical deployment scenario; regardless, using the IP-ID as a channel to convey a HOST_ID is ill-advised.

An alternate way to convey the HOST_ID is to define an IP option . A HOST_ID IP option can be inserted by the address-sharing function to uniquely distinguish a host among those sharing the same IP address. An example of such an option is documented in . This IP option allows the conveyance of an IPv4 address, an IPv6 prefix, a Generic Routing Encapsulation (GRE) key, an IPv6 Flow Label, etc. An IP option may also be used as described in Section 4.6 of .

This proposal can apply to any transport protocol. However, it is widely known that routers and other middleboxes filter IP options (e.g., drop IP packets with unknown IP options, strip unknown IP options, etc.). Injecting the HOST_ID IP option introduces some implementation complexity in the following cases: The packet is at or close to the MTU size. The options space is exhausted. Previous studies demonstrated that "IP Options are not an option" (refer to and ). In conclusion, using an IP option to convey a HOST_ID is not viable.

The HOST_ID may be conveyed in a dedicated TCP option. An example is specified in . This option encloses the TCP client's identifier (e.g., the lower 16 bits of its IPv4 address, its VLAN ID, VRF ID, or subscriber ID). The address-sharing device inserts this TCP option into the TCP SYN packet.

Using a new TCP option to convey the HOST_ID does not require any modification to the applications, but it is applicable only for TCP&nbhy;based applications. Applications relying on other transport protocols are therefore left unsolved. discusses the interference with other TCP options. The risk of session failure due to handling a new TCP option is low as measured in . provides a detailed implementation and experimentation report of a HOST_ID TCP option. This document provides an in-depth investigation of the impact of implementing HOST_ID on the host, the address-sharing function, and the enforcement of policies at the server side. It also reports a failure ratio of 0.103% among the top 100,000 websites. Some downsides have been identified with defining a TCP option to reveal a host identity: Conveying an IP address in a TCP option may be seen as a violation of OSI layers, but since IP addresses are already used for the checksum computation, this is not seen as a blocking point. Moreover, the updated version of no longer allows conveyance of a full IP address because the HOST_ID is encoded in 16 bits. TCP option space is limited and might be consumed by the TCP client. discusses two approaches to sending the HOST_ID: sending the HOST_ID in the TCP SYN (which consumes more bytes in the TCP header of the TCP SYN) and sending the HOST_ID in a TCP ACK (which consumes only two bytes in the TCP SYN). Content providers may find it more desirable to receive the HOST_ID in the TCP SYN, as that more closely preserves the HOST_ID received in the source IP address as per current practices. Moreover, sending the HOST_ID in the TCP SYN does not interfere with . In the ACK mode, if the server is configured to deliver different data based on HOST_ID, then it would have to wait for the ACK before transmitting data. HOST_ID mechanisms need to be aware of end-to-end (E2E) issues and avoid interfering with them. One example of such interference would be injecting or removing TCP options of transited packets; another such interference involves terminating and re-originating TCP connections not belonging to the transit device. The HOST_ID TCP option handled by the source node avoids this issue. Injecting the HOST_ID TCP option introduces some implementation complexity if the options space is exhausted. Specification document(s) should specify the behavior of the address-sharing function in detail in such a case. It is more complicated to implement sending the HOST_ID in a TCP ACK, as it can introduce MTU issues if the ACK packet also contains TCP data or if a TCP segment is lost. Note that MTU complications can be experienced if user data is included in a SYN packet (e.g., ). When there are several NATs in the path, the original HOST_ID may be lost. The loss of the original HOST_ID may not be a problem, as the target usage is between proxies or between a CGN and server. Only the information leaked in the last communication leg (i.e., between the last address-sharing function and the server) is likely to be useful. Interference with usages such as a Forwarded HTTP header (see ) should be elaborated to specify the behavior of servers when both options are used; in particular, specify which information to use: the content of the TCP option or what is conveyed in the application headers. When load balancers or proxies are in the path, this option does not allow the preservation of the original source IP address and source port. Preserving such information is required for logging purposes (e.g., ). defines a TCP option that allows various combinations of source information (e.g., source port, source port and source IP address, source IPv6 prefix, etc.) to be revealed. More discussion about issues raised when extending TCP can be found at .

Another option is to not require any change within the transport or the IP levels but to convey the required information that will be used to disambiguate hosts at the application payload. The format of the conveyed information and the related semantics depend on its application (e.g., HTTP, SIP, SMTP, etc.). Related mechanisms could be developed for other application-layer protocols, but the discussion in this document is limited to HTTP and similar protocols. For HTTP, the Forwarded header can be used to display the original IP address when an address-sharing device is involved. Service providers operating address-sharing devices can enable the feature of injecting the Forwarded header, which will enclose the original IPv4 address or the IPv6 prefix part (see the example shown in ). The address-sharing device has to strip all included Forwarded headers before injecting its own. Servers may rely on the contents of this field to enforce some policies such as blacklisting misbehaving users. Note that standardizes the Forwarded header field, to replace the de facto (and not standard) X-Forwarded-For (XFF) header.

Not all applications impacted by address sharing can support the ability to disclose the original IP address. Only a subset of protocols (e.g., HTTP) can rely on this solution. For the HTTP case, to prevent users from injecting invalid HOST_IDs, an initiative has been launched by Wikimedia to maintain a list of trusted ISPs (Internet Service Providers) using XFF (see the list available at ). If an address-sharing device is on the list of trusted XFF ISPs, users editing Wikimedia located behind the address-sharing device will appear to be editing from their "original" IP address and not from the NATed IP address. If an offending activity is detected, individual hosts can be blacklisted instead of blacklisting all hosts sharing the same IP address. XFF header injection is a common practice of load balancers. When a load balancer is in the path, the original content of any included XFF header should not be stripped. Otherwise, the information about the "origin" IP address will be lost. When several address-sharing devices are crossed, the Forwarded header can convey the list of IP addresses (e.g., ). The origin HOST_ID can be exposed to the target server. Injecting the Forwarded header also introduces some implementation complexity if the HTTP message is at or close to the MTU size. It has been reported that some HTTP proxy implementations may encounter parsing issues when injecting an XFF header. Injecting the Forwarded header for all HTTPS traffic is infeasible. This may be problematic given the current HTTPS usage trends.

The solution, referred to as the Proxy Protocol , does not require any application-specific knowledge. The proposed solution (Proxy Protocol Version 1) would insert identification data directly into the application-data stream prior to the actual protocol data being sent, regardless of the protocol. Every application protocol would begin with a textual string of "PROXY", followed by some textual identification data, and with a CRLF; only then would the application data be inserted. shows an example of a line of data used for this purpose, in this case, for a TCP-over-IPv4 connection received from 192.0.2.1:56324 and destined to 192.0.2.15:443.

Upon receipt of a message conveying this line, the server removes the line from the incoming stream. The line is parsed to retrieve the transported protocol. The content of this line is recorded in logs and used to enforce policies. Proxy Protocol Version 2 is designed to accommodate IPv4/IPv6 and also non-TCP protocols (see for more details).

This solution can be deployed in a controlled environment, but it cannot be deployed to all access services available in the Internet. If the remote server does not support the Proxy Protocol, the session will fail. Other complications will arise due to the presence of firewalls, for instance. As a consequence, this solution is infeasible and cannot be recommended.

This solution does not require any action from the address-sharing function to disclose a host identifier. Instead of assuming that all transport ports are associated with one single host, each host under the same external IP address is assigned a restricted port set. These port sets are then advertised to remote servers using offline means. This announcement is not required for the delivery of internal services (i.e., offered by the service provider deploying the address-sharing function) relying on implicit identification. Port sets assigned to hosts may be static or dynamic. Port set announcements to remote servers are not required to reveal the identity of individual hosts; they are used to advertise the enforced policy to generate non-overlapping port sets (e.g., the transport space associated with an IP address is fragmented to contiguous blocks of 2048 port numbers). Examples of such an approach are documented in and .

The solution does not require defining new fields or options; it is policy based. The solution may contradict the port randomization as identified in . A mitigation would be to avoid assigning static port sets to individual hosts. The method is convenient for the delivery of services offered by the service provider that is also offering the Internet access service.

specifies an architecture that introduces a new namespace to convey identity information.

This solution requires both the client and the server to support HIP . Additional architectural considerations are to be taken into account, such as the key exchanges. An alternative deployment model that does not require the client to be HIP-enabled is having the address-sharing function behave as a UDP/TCP-HIP relay. This model is also not viable as it assumes all servers are HIP-enabled. This solution is a theoretical construct (i.e., the proposal is not documented).

Another alternative is to convey the HOST_ID using a separate notification channel than the one the packets issued to invoke the service. This solution relies on a mechanism where the address-sharing function encapsulates the necessary host-identifying information into an ICMP Echo Request packet that it sends in parallel with the initial session creation (e.g., SYN). The information included in the ICMP Request Data portion describes the five-tuples as seen on both sides of the address-sharing function. An implementation example is defined in .

This ICMP proposal is valid for any transport protocol that uses a port number. The address-sharing function may be configured with the transport protocols that will trigger issuing those ICMP messages. A hint should be provided to the ultimate server (or intermediate nodes) that the ICMP Echo Request conveys a HOST_ID. This may be implemented using magic numbers (a magic number is a self-selected codepoint whose primary value is its unlikely collision with values selected by others). Even if ICMP packets are blocked in the communication path, the user connection does not have to be impacted. Implementations requiring a session establishment to be delayed until receipt of the companion ICMP Echo Request may lead to some user-experience degradation. Because of the presence of load balancers in the path, the ultimate server receiving the SYN packet may not be the one that receives the ICMP message conveying the HOST_ID. Because of the presence of load balancers in the path, the port number assigned by address sharing may be lost. Therefore, the mapping information conveyed in the ICMP may not be sufficient to associate a SYN packet with a received ICMP. The proposal is not compatible with the presence of cascaded NAT. The main reason is that each NAT in the path will generate an ICMP message to reveal the internal host identifier. Because these messages will be translated by the downstream address-sharing devices, the remote server will receive multiple ICMP messages and will need to decide which host identifier to use. The ICMP proposal will add traffic overhead for both the server and the address-sharing device. The ICMP proposal is similar to other mechanisms (e.g., IPFIX and Syslog ) for reporting dynamic mappings to a mediation platform (mainly for legal traceability purposes). Performance degradation is likely to be experienced by address-sharing functions because ICMP messages are sent for each new instantiated mapping (even if the mapping exists). In some scenarios (e.g., Section 3 of ), the HOST_ID should be interpreted by intermediate devices that embed Policy Enforcement Points (PEP) responsible for granting access to some services. These PEPs need to inspect all received packets in order to find the companion (traffic) messages to be correlated with ICMP messages conveying HOST_IDs. This induces more complexity to these intermediate devices.

Another alternative is to retrieve the HOST_ID using a dedicated query channel. An implementation example may rely on the Identification Protocol (Ident) . This solution assumes that the address-sharing function implements the server part of IDENT, while remote servers implement the client part of the protocol. IDENT needs to be updated to be able to return a host identifier instead of the user-id as defined in . The IDENT response syntax uses the same USERID field described in , but rather than returning a username, a host identifier (e.g., a 16-bit value) is returned. For any new incoming connection, the server contacts the IDENT server to retrieve the associated identifier. During that phase, the connection may be delayed.

IDENT is specific to TCP. Alternative out-of-band mechanisms may be designed to cover other transport protocols such as UDP. This solution requires the address-sharing function to embed an IDENT server. A hint should be provided to the ultimate server (or intermediate nodes) that the address-sharing function implements the IDENT protocol, for example, publishing this capability using DNS (other solutions can be envisaged). An out-of-band mechanism may require some administrative setup (e.g., contract agreement) between the entity managing the address-sharing function and the entity managing the remote server. Such a deployment is not feasible in the Internet at large because establishing and maintaining agreements between ISPs and all service actors is burdensome and not scalable. Implementations requiring delay of the establishment of a session until receipt of the companion IDENT response may lead to some user-experience degradation. The IDENT proposal will add traffic overhead for both the server and the address-sharing device. Performance degradation is likely to be experienced by address-sharing functions embedding the IDENT server. This is further exacerbated if the address-sharing function has to handle an IDENT query for each new instantiated mapping (even if the mapping exists). In some scenarios (e.g., Section 3 of ), the HOST_ID should be interpreted by intermediate devices that embed Policy Enforcement Points (PEP) responsible for granting access to some services. These PEPs need to inspect all received packets in order to generate the companion IDENT queries. This may induce more complexity to these intermediate devices. IDENT queries may be generated by illegitimate TCP servers. This would require the address-sharing function to enforce some policies (e.g., rate-limit queries, filter based on the source IP address, etc.).

Table 1 summarizes the approaches analyzed in this document. "Encrypted Traffic" refers to TLS. The use of IPsec and its complications traversing NATs are discussed in Section 2.2 of . Similar to what is suggested in Section 13.5 of , HOST_ID specification document(s) should analyze the compatibility of each IPsec mode in detail. "Success ratio" indicates the ratio of successful communications with remote servers when the HOST_ID is injected using a proposed solution. More details are provided below to explain how the success ratio is computed for each proposed solution. "Possible Perf Impact" indicates the level of expected performance degradation. The indicated degradation is an estimate based on the need for processing at the IP layer. "OS TCP/IP Modif" indicates whether a modification of the OS TCP/IP stack is required at the server side. "Deployable today" indicates if the solution can be generalized without any constraint on current architectures and practices.

Provided success ratio figures for TCP and IP options are based on the results documented in and . The provided success ratio for the IP-ID is theoretical; it assumes the address-sharing function follows the rules (see )to rewrite the IP Identification field. Since PROXY and HIP are not widely deployed, the success ratio for establishing communication with remote servers using these protocols is low. The success ratio for the ICMP-based solution is implementation-specific but it is likely to be close to 100%. The success ratio depends on how efficiently the solution is implemented on the server side. A remote server that does not support the ICMP-based solution will ignore received companion ICMP messages. An upgraded server will need to delay the acceptance of a session until it receives the companion ICMP message. The success ratio for the IDENT solution is implementation specific but it is likely to be close to 100%. The success ratio depends on how efficient the solution is implemented on the server side. A remote server that does not support IDENT will accept a session establishment request following its normal operation. An upgraded server will need to delay the acceptance of a session until it receives a response to the IDENT request it will send to the host. The provided success ratio for the Port Set and HTTP header solutions is 100% because no additional Layer 3 or Layer 4 field is needed to convey HOST_ID for these solutions.

If the server trusts the content of the HOST_ID field, a third-party user can be impacted by a misbehaving user revealing a "faked" HOST_ID (e.g., original IP address). This same security concern applies for the injection of an IP option, TCP option, and application-related content (e.g., the Forwarded HTTP header) by the address-sharing device. The HOST_ID may be used to leak information about the internal structure of a network behind an address-sharing function. If this behavior is undesired for the network administrator, the address-sharing function can be configured to strip any existing HOST_ID in received packets from internal hosts. HOST_ID specification documents should elaborate further on threats inherent to each individual solution used to convey the HOST_ID (e.g., use of the IP-ID field to count hosts behind a NAT ). For more discussion of privacy issues related to the HOST_ID, see .

Many thanks to D. Wing, C. Jacquenet, J. Halpern, B. Haberman, and P. Yee for their review, comments, and inputs. Thanks also to P McCann, T. Tsou, Z. Dong, B. Briscoe, T. Taylor, M. Blanchet, D. Wing, and A. Yourtchenko for the discussions in Prague. Some of the issues related to defining a new TCP option have been raised by L. Eggert. The privacy text was provided by A. Cooper.