<?xml version='1.0' encoding='utf-8'?>
<?xml-stylesheet type="text/xsl" href="rfc2629.xslt" ?>
<!-- generated by https://github.com/cabo/kramdown-rfc2629 version 1.3.3 -->
<!DOCTYPE rfc SYSTEM "rfc2629-xhtml.ent">
<?rfc toc="yes"?>
<?rfc sortrefs="yes"?>
<?rfc symrefs="yes"?>
<rfc xmlns:xi="http://www.w3.org/2001/XInclude" ipr="trust200902" docName="draft-irtf-pearg-website-fingerprinting-01" category="info" obsoletes="" updates="" submissionType="IETF" xml:lang="en" tocInclude="true" sortRefs="true" symRefs="true" version="3">
  <!-- xml2rfc v2v3 conversion 2.42.0 -->
  <front>
    <title abbrev="Network-Based Website Fingerprinting">Network-Based Website Fingerprinting</title>
    <seriesInfo name="Internet-Draft" value="draft-irtf-pearg-website-fingerprinting-01"/>
    <author initials="I." surname="Goldberg" fullname="Ian Goldberg">
      <organization>University of Waterloo</organization>
      <address>
        <email>iang@uwaterloo.ca</email>
      </address>
    </author>
    <author initials="T." surname="Wang" fullname="Tao Wang">
      <organization>HK University of Science and Technology</organization>
      <address>
        <email>taow@cse.ust.hk</email>
      </address>
    </author>
    <author initials="C.A." surname="Wood" fullname="Christopher A. Wood">
      <organization>Cloudflare</organization>
      <address>
        <email>caw@heapingbits.net</email>
      </address>
    </author>
    <date year="2020" month="September" day="08"/>
    <area>General</area>
    <workgroup>pearg</workgroup>
    <keyword>Internet-Draft</keyword>
    <abstract>
      <t>The IETF is well on its way to protecting connection metadata with protocols such as DNS-over-TLS
and DNS-over-HTTPS, and work-in-progress towards encrypting the TLS SNI. However, more
work is needed to protect traffic metadata, especially in the context of web traffic.
In this document, we survey Website Fingerprinting attacks, which are a class of attacks that
use machine learning techniques to attack web privacy, and highlight metadata leaks used
by said attacks. We also survey proposed mitigations for such leakage and discuss their
applicability to IETF protocols such as TLS, QUIC, and HTTP. We endeavor to show that
Website Fingerprinting attacks are a serious problem that affect all Internet users,
and we pose open problems and directions for future research in this area.</t>
    </abstract>
    <note>
      <name>Note to Readers</name>
      <t>Source for this draft and an issue tracker can be found at
  <eref target="https://github.com/chris-wood/ietf-fingerprinting">https://github.com/chris-wood/ietf-fingerprinting</eref>.</t>
    </note>
  </front>
  <middle>
    <section anchor="introduction" numbered="true" toc="default">
      <name>Introduction</name>
      <t>Internet protocols such as TLS 1.3 <xref target="RFC8446" format="default"/> and QUIC <xref target="I-D.ietf-quic-transport" format="default"/>
bring substantial improvements to end-users.
The IETF engineered these with security and privacy in mind by encrypting
more protocol messages using modern cryptographic primitives and algorithms, and engineering
against flaws found in previous protocols, yielding several desirable security
properties, including: forward-secure session key secrecy, downgrade protection, key
compromise impersonation resistance, and protection of endpoint identities.
Combined, these two protocols are set to protect a significant amount of Internet data.
However, significant metadata leaks still exist for users of these protocols. Examples include
plaintext TLS SNI and application-specific extensions (ALPN), as well as DNS queries. This information
can be used by a passive attacker to learn information about the contents of an otherwise
encrypted network connection. Recently, such information has also been studied as a means
of building unique user profiles <xref target="li2018can" format="default"/>. It has also been used to build flow classifiers
that aid network management <xref target="foremski2014dns" format="default"/>.</t>
      <t>In the context of Tor, a popular low-latency anonymity network, a common class of attacks that use
metadata for such inference is called Website Fingerprinting (WF).
These attacks use machine learning techniques built with
features extracted from metadata such as traffic patterns to attack web (browsing) privacy.
Miller et al. <xref target="miller2014know" format="default"/> show how these attacks can be applied to web browsing
traffic protected with HTTPS to reveal private information about users.
Pironti et al. <xref target="pironti2012identifying" format="default"/> use similar attacks based on data sizes to
identify individual social media clients using encrypted connections.
Fingerprinting attacks using encrypted traffic analysis are also applicable to encrypted
media streams, such as Netflix videos. (See work from Reed et al. <xref target="reed2017identifying" format="default"/>
and Schuster et al. <xref target="schuster2017beauty" format="default"/>
for examples of these attacks.) WF attacks have also been applied to other IETF
protocols such as encrypted DNS, including dnscrypt, DNS-over-TLS,
and DNS-over-HTTPS <xref target="siby2018dns" format="default"/><xref target="shulman2014pretty" format="default"/>. In the past, they have also
been conducted remotely <xref target="gong2010fingerprinting" format="default"/>, using buffer-based side channels
in a victim's home router.</t>
      <t>Protocols such as DNS-over-TLS and DNS-over-HTTPS <xref target="RFC8484" format="default"/>, and work-in-progress
towards encrypting the TLS SNI extension <xref target="I-D.ietf-tls-esni" format="default"/>, help minimize metadata
sent in cleartext on the wire. However, regardless of protocol and even network-layer
fingerprinting mitigations, application layer specifics, e.g., web page sizes and client
request patterns, reveal a noticeable amount of information to attackers. We argue
that much more work is needed to protect encrypted connection metadata, especially
in the context of web traffic.</t>
      <t>In this document, we describe WF attacks in the context of IETF protocols such as TLS and
QUIC. We survey WF attacks and highlight metadata features and classification techniques
used to conduct said attacks. We also describe proposed mitigations for these attacks
and discuss their applicability to IETF protocols. We conclude with a discussion of open
problems and directions for future research and advocate for more work in this area.</t>
    </section>
    <section anchor="background" numbered="true" toc="default">
      <name>Background</name>
      <t>In this section we review how most secure Internet connections are made today. We omit custom
configurations such as those using VPNs and proxies since they do not represent the common case
for most Internet users. The following steps briefly describe the sequence of events that
normally occur when a web client, e.g., browser, curl, etc., connects to a website and obtains
some resource. First an unencrypted DNS query is sent to an untrusted DNS recursive
resolver to resolve a name to an IP address. Upon receipt, clients then open a TCP and TLS
connection to the destination address. During this stage, metadata such as the TLS SNI and ALPN
values are sent in cleartext. The SNI is used to denote the destination application or endpoint
to which clients want to connect. Servers use this for several purposes, including selecting
an appropriate certificate (one with the SNI name in the SubjectAlternativeName list) or
routing to a different backend terminator. ALPN values are used to negotiate which application-layer
protocol will be used on top of the TLS connection. Common values include "http/1.1", "h2", and
(soon) "h3". Upon connection, clients then send HTTP messages to obtain the desired resource.</t>
      <!-- figure 1 here -->

<t>Connections look different (on the wire) with TLS 1.3, encrypted DNS via DNS-over-TLS or
DNS-over-HTTPS, and encrypted SNI.
DNS queries are encrypted to a (trusted) recursive resolver and TLS metadata such as SNI
are encrypted in transit to the terminator. Despite the reduction in cleartext metadata
sent over the wire, there still remains several sources of information that an adversary
may use for malicious purposes, including: size and timing of DNS queries and responses,
size and timing or application traffic, and connection attempts induced while loading a
web resource, e.g., Javascript files. So while technologies such as Encrypted SNI, DoT,
and DoH help protect some metadata, they are not complete solutions to the larger problem.
In the following section, we discuss this overarching problem in detail.</t>
      <!-- figure 2 here -->

</section>
    <section anchor="website-fingerprinting" numbered="true" toc="default">
      <name>Website Fingerprinting</name>
      <t>Website Fingerprinting (WF) is a class of attacks that exploit metadata leakage to attack
end-user privacy on the Internet. In the WF threat model,
Adv is assumed to be a passive and local attacker. Local means that Adv can associate
traffic with a given client. Examples include proxies to which clients directly connect.
Passive means that Adv can only view traffic in transit. It cannot add, drop, or otherwise
modify packets between the victim client and server(s). Use of reliable and encrypted transport
protocols such as TLS limit on-path attackers to eavesdropping on encrypted packets. (In
QUIC, however, reordering packets is possible.)</t>
      <t>Traffic features used for classification include properties such as packet size, timing,
direction, interarrival times, and burstiness, among many others <xref target="wang2016website" format="default"/>. Normally, features
are restricted to those which are extractable as a passive eavesdropper, and not those which
are viewable by modifying client or server behavior. Specifically, this means that
attacks such as CRIME <xref target="CRIME" format="default"/> and TIME <xref target="TIME" format="default"/>, which rely on an attacker abusing TLS-layer compression
to leak contents of an encrypted connection, are out of scope.</t>
      <t>Website Fingerprinting attacks have evolved over the years through three phases:
(1) Closed-world WF on SSL/TLS, (2) Closed-world WF on Tor, and (3) Open-world WF on Tor.</t>
      <ol spacing="normal" type="1">
        <li>In the closed-world model, clients are assumed to only visit a small set of pages monitored by Adv.
This is less realistic but easier to analyze than the open-world model discussed below,
and so the earliest results achieved success on SSL/TLS in this model.
(For a realistic attack, Adv would need to monitor every possible page of interest to each client, which is impractical.)
Attacks against proxy-based privacy technologies such as VPNs and SSH tunneling,
which has almost no effect on the network, falls under this category as well.</li>
        <li>Tor, an anonymity network built on onion routing, is harder to attack than SSL for several reasons;
successful results on Tor thus came later.
First, Tor pads all cells (Tor's application-layer datagrams) to the same constant size,
removing unique packet lengths as a powerful feature for the attacker.
Second, Tor imposes random network conditions upon the client due to random selection of proxies,
so packet sequences are less likely to be consistent.</li>
        <li>In the open-world model, Adv wishes to learn whenever a victim
client visits one of a select number of monitored pages <xref target="wang2016website" format="default"/>. Adversaries train
classifiers in this model using monitored and non-monitored websites of their choosing. By
definition, Adv cannot train using client-chosen pages. Clients then visit pages at will
and Adv attempts to learn whenever a monitored page is visited, if any are at all.
This is a realistic model capturing the fact that the set of pages any attacker would be interested in must
necessarily be a small subset of the set of all pages.
As this is a harder model to attack, successful results on this model came later.</li>
      </ol>
    </section>
    <section anchor="attacks" numbered="true" toc="default">
      <name>Attacks</name>
      <ol spacing="normal" type="1">
        <li>Closed-world WF on TLS: WF attacks date back to applications on SSL first inspired by Wagner and
Schneier <xref target="wagner1996analysis" format="default"/>, in which the authors observed that
packet lengths reveal information about the underlying data. Subsequent attacks
carried out by Cheng et al. <xref target="cheng1998traffic" format="default"/>,
Sun et al. <xref target="sun2002statistical" format="default"/>, and Hintz <xref target="hintz2002fingerprinting" format="default"/> continued to show access.
These attacks assume Adv has knowledge of the target resource length(s),
which is not always possible with techniques such as padding.</li>
      </ol>
      <t>Bissias et al. <xref target="bissias2005privacy" format="default"/> use cross correlation of inter-packet
times in one second time windows as an WF attack. Danezis <xref target="danezis2009traffic" format="default"/> model
websites using a Hidden Markov Model (HMM) and use it, along with TLS traffic traces
revealing only approximate lengths, to identify requested resources on a page. Their results
vary the amount of information available to an adversary when building the HMM. Even in cases
where resource popularity is omitted, which reflects the case where an adversary scrapes static
websites, resource recall was high (86\%).
Liberatore and Levine <xref target="liberatore2006inferring" format="default"/> proposed two WF attacks
using the Jaccard coefficient and the Naive Bayes classifier.
Herrmann et al. <xref target="herrmann2009website" format="default"/> extended the work of Liberatore and Levine with a multinomial
Naive Bayes classifier computed using three input frequency transformations.
Results yielded higher accuracy than that of Liberatore and Levine.
Herrmann's attack is the best in this category, but the authors assume packets which do not fill a MTU
represent packet trailers. Therefore, uniqueness is only accurate modulo the MTU. Efficacy
is limited if endpoints pad packets to the MTU or another fixed length. Modern protocols
such as HTTP/2, QUIC, and TLS 1.3 all provide some form of application-controlled padding.
(Note: These attacks are not successful on Tor.)</t>
      <ol spacing="normal" type="1">
        <li>Closed-world WF on Tor: Shmatikov and Wang <xref target="shmatikov2006timing" format="default"/>
presented a WF attack that exploits cross correlation of arrival packet counts in
one second time windows.
Lu et al. <xref target="lu2010website" format="default"/> developed a classifier based on the Levenshtein distance between
ingress and egress packet lengths extracted from packet sequences. Distance is computed between
strings of ingress and egress packet lengths. The training packet sequence with the closest
distance to the testing packet sequence is deemed the match. Dyer et al. <xref target="dyer2012peek" format="default"/>
used a Naive Bayes classifier trained with a reduced set of features, including total
response transmission time, length of packets (in each direction), and burst lengths. (Wang
<xref target="wang2016website" format="default"/> notes that measuring burst lengths in Tor is difficult given the presence of
SENDME cells for flow control.) This approach did not yield any measurable improvements over
the SVM classifier from Panchenko et al. Cai et al. <xref target="cai2012touching" format="default"/> extend the work of Lu et al.
by adding transpositions to the Levenshtein distance computation and normalizing the result,
yielding what the authors refer to as the Optimal String Alignment Distance (OSAD).
Before feature extraction, the authors round TCP packet lengths to the nearest multiple of
600B as an estimate of the number of Tor cells.</li>
      </ol>
      <t>Wang et al. <xref target="wang2013improved" format="default"/> tuned the OSAD-based attack to improve its accuracy. Specific changes
include use of Tor cells instead of TCP packets for packet and burst lengths, as well as heuristics
to remove SENDME cells (those not carrying application data) from flows to recover true
burst lengths. The authors also modified the distance computation by removing substitutions,
increasing the weight for egress packets, and varying the transposition cost across the packet
sequence (large weights at the beginning of a trace, and smaller weights near the end, where
variations are expected across repeated page loads.) Wang et al. also developed an alternate classifier
with lower accuracy yet superior performance (quadratic to linear time complexity). It works by
minimizing the sum of two costs: sequence transpositions and sequence deletions or insertions. These
two costs are computed separately, in contrast to the first approach which computes them simultaneously.</t>
      <t>Hayes et al. <xref target="hayes2016k" format="default"/> developed an attack called k-fingerprinting, which uses a k-NN classifier
with features ranked by random decision forests. Their feature set includes timing information, e.g.,
statistics on packets per second, among the higher ranked features. (Higher ranked features have more
weight in the classification phase.) Yan et al. <xref target="yan2018feature" format="default"/> used similar (manually curated)
features with a CNN-based classifier. Time-based features were among the more effective features
identified. Rahman et al. <xref target="rahman2019tik" format="default"/> improved time-based features by focusing on bursts,
e.g., burst length, variance, inter-burst delay, etc., rather than more granular per-packet statistics.
(The latter tend to vary for inconsistencies across packet traces for websites.) This improved accuracy
of existing Deep Learning attacks from Sirinam et al. <xref target="sirinam2018deep" format="default"/>, especially when coupled
with packet direction information.</t>
      <ol spacing="normal" type="1">
        <li>Open-world WF on Tor and TLS: Panchenko et al. <xref target="panchenko2011website" format="default"/>
were the first to use a support vector machine (SVM) classifier trained with web domain-specific
features, such as HTML document sizes, as well as packet lengths.
Wang et al. <xref target="wang2014effective" format="default"/> also developed an attack using a k-Nearest Neighbors (k-NN) classifier,
which is a supervised machine learning algorithm, targeting the open world setting. The classifier
extracts a large number of features from packet sequences, including raw (ingress and egress)
packet counts, unique packet lengths, direction, burst lengths, and inter-packet times, among others.
(There are 4226 features in total.) The k-NN distance metric is computed as the sum of weighted
feature differences.</li>
      </ol>
      <t>Kota et al. <xref target="abe2016fingerprinting" format="default"/> were the first to use Deep Learning (DL) methods based on Stacked
Denoising Autoencoders for WF attacks. (Autoencoders reduce feature input dimensions when stacked.)
Kota et al. form input vectors from Tor cell directions (+1 or -1). They use no other features.
Using a (small) data set from Wang <xref target="wang2016website" format="default"/>, the classifier achieves a 86% true positive
rate and 2% false positive rate in the open world model. Rimmer et al. <xref target="rimmer2018automated" format="default"/>
applied DL for automated feature generation and classifier construction. Trained with 2,500 traces per
website, their system achieves 96.3% accuracy in the open world model.
Recently, Bhat et al. <xref target="bhat2018var" format="default"/>, Oh et al. <xref target="oh2017pfp" format="default"/>, and Sirinam et al. <xref target="sirinam2018deep" format="default"/>
used Convolutional Neural Networks (CNNs) and Deep Neural Networks (DNNs) for WF attacks. Results from
Sirinam et al. show the best results - 98% on Tor without recent defenses (in Section {{defenses}) -
while performing favorably when select defenses are used for both open and closed world models.</t>
      <t>Yan et al. <xref target="yan2018feature" format="default"/> studied manual high-information feature extraction from packet traces.
They "exhaustively" examined different levels of features, including packet, burst, TCP, port, and IP address,
summing to 35,683 in total, and distilled them into a diverse set of uncorrelated features for eight
different communication scenarios. Rahman <xref target="rahman2018using" format="default"/> studied the utility of features derived
from packet interarrival times, including: median interarrival time (per burst), burst packet arrival
time variance, cross-burst interarrival median differences, and others. Using a CNN, results show that
these features yield a non-negligible increase in WF attack accuracy.</t>
      <!-- end of description -->

</section>
    <section anchor="base-rate-fallacy" numbered="true" toc="default">
      <name>Base Rate Fallacy</name>
      <t>For all WF attacks, one limitation worth highlighting is the base rate fallacy. This can be summarized
as follows: highly accurate classifiers with a reliable false positive rate (FPR) decrease in
efficacy as the world size increases. Juarez et al. <xref target="juarez2014critical" format="default"/> studied its impact by
measuring the Bayesian detection rate (BDR) in comparison to the FPR as a function of world size.
As the world size increases, the BDR approaches 0 while the FPR remains stable, meaning that the probability
of incorrect classifier results increase as well. Juarez et al. partially address the base rate fallacy
problem by adding a confirmation step to their classifier.
Another problem is that web content is (increasingly) dynamic. Most WF attacks, especially those in closed
world models, assume that traces are static. However, Juarez et al. <xref target="juarez2014critical" format="default"/> show
this is not the case even for "simple" pages such as google.com. Thus, due to the base fallacy
rate and dynamic nature of content, classifiers require continual retraining in order to ensure accuracy.</t>
    </section>
    <section anchor="defenses" numbered="true" toc="default">
      <name>Defenses</name>
      <t>There are at least two types of WF defenses: traffic shaping or morphing algorithms, and traffic
splitting algorithms. This section describes and illustrates examples of both.</t>
      <section anchor="traffic-morphing" numbered="true" toc="default">
        <name>Traffic Morphing</name>
        <t>WF defenses are deterministic or randomized algorithms that take as input application data or packet sequences
and return modified application data or packet sequences. Viable defenses seek to minimize the transformation
cost and maximum (theoretical and perfect) attacker accuracy. Naive defenses such as sending a constant stream
of (possibly random) bytes between client and server may be effective though clearly not viable from a cost
perspective. Relevant cost metrics include bandwidth overhead, added time or latency (and its impact on related
metrics such as page load time), and even CPU cost, though the latter is often ignored in favor of the former two.
Wang <xref target="wang2016website" format="default"/> describe defenses as either limited or general. A limited defense is
one which only helps mitigate specific WF attacks by transforming packets in a way to obviate a particular
(set of) feature(s) used by said attacks. In contrast, general defenses help mitigate a variety of attacks.</t>
        <t>Several general defenses have been proposed, including BuFLO <xref target="dyer2012peek" format="default"/>, which pads packets to
a fixed length of 1500B (the normal MTU) and schedules packets for transmission at fixed period intervals
(and sends fake data if nothing is yet available). Tamaraw <xref target="wang2014comparing" format="default"/> is an improvement over BuFLO
that uses two different fixed lengths for packet transmission, rather than one, to save on bandwidth overhead.
Tamaraw also uses two different scheduling rates for ingress and egress packets. The authors chose to make
the ingress packet period smaller than the egress packet period since HTTP responses are often larger in size
and count - if HTTP Push is used - than requests. While provably correct, both BuFLO and Tamaraw limit
the rate at which clients send traffic, and requires all clients to send at a uniform rate. Both requirements
therefore make it difficult to apply as a generic defense in IETF protocols.</t>
        <t>Wang et al. also developed Supersequence <xref target="wang2016website" format="default"/>, which attempts to approximate a bandwidth-optimal
deterministic defense. This is done by casting the padding and flow control problem as the shortest common
subsequence (SCS) of the transformed packet trace. Supersequence approximates the solution by learning the optimal
packet scheduling rate; it uses the same padding scheme as Tamaraw.</t>
        <t>Walkie-Talkie <xref target="wang2015walkie" format="default"/> is a collection of mechanisms for WF defense. It includes running
the client (browser) in half-duplex mode to batch requests and responses together, as well as randomly
padding traffic so as to mimic traffic of benign websites. It assumes knowledge of traffic patterns for
benign websites, which can be information learned over time or provided by a cooperating peer. Goldberg
and Wang also propose a "randomized" variant that pads real bursts of requests and generates random
request bursts according to a uniform distribution. The half-duplex mode could be implemented as an extension
to a protocol such as HTTP/2, QUIC, or TLS.</t>
        <t>Many limited defenses have also been proposed. We list prominent works below.</t>
        <ul spacing="normal">
          <li>Shmatikov and Wang <xref target="shmatikov2006timing" format="default"/> developed adaptive padding which adds packets to mask
inter-packet times. (This mechanism does not ever delay application data being sent, in contrast to other
padding mechanisms such as BuFLO; see below.)
Juarez et al. <xref target="juarez2015wtf" format="default"/>}<xref target="juarez2016toward" format="default"/> also created a WF defense based on adaptive padding called WTF-PAD.
This variant uses application data and "gap" distribution to generate padding for delays. Specifically, when
not sending application data, senders use the gap distribution to drive fake packet transmission.
WTF-PAD can be run by a single endpoint, though it is assumed that both client and server participate.
As mentioned above, protocols such as HTTP/2, QUIC, and TLS 1.3 offer a mechanism by which applications can
send padding. WTF-PAD could therefore be implemented as an extension to any of these protocols, either by
applications supplying padding distributions or the system learning them over time.</li>
          <li>In the context of HTTP, Danezis <xref target="danezis2009traffic" format="default"/> proposed padding: URLs, content, and even HTML
document structures to mask application data lengths.</li>
          <li>Wright et al. <xref target="wright2009traffic" format="default"/> developed traffic morphing, which pads packets in such a way
so as to make the sequence from one page have characteristics of another (non-monitored or benign) page.
This technique requires application-specific knowledge about benign pages and is therefore best implemented
outside of the transport layer.</li>
          <li>Nithyanand et al. <xref target="nithyanand2014glove" format="default"/> developed a mechanism called Glove,
wherein traces were first clustered and then morphed (via dummy insertion, packet
merging, splitting, and delaying) to look indistinguishable within clusters. When used
to protect the Alexa top 500 domains, Glove performs well with respect to bandwidth
overhead when compared to BuFLO and CS-BuFLO. Varying the cluster size can tune
Glove's bandwidth overhead.</li>
          <li>Pironti et al. <xref target="pironti2012identifying" format="default"/> developed a TLS-based fragmentation and padding
scheme designed to hide the length of application data within a certain range with record padding.
The mechanism works by iteratively splitting application data into variable sized segments. Applications
can guide the range of viable lengths provided such information is available.</li>
          <li>Luo et al. <xref target="luo2011httpos" format="default"/> created HTTPS with Obfuscation (HTTPOS), which is a client-side
mechanism for obfuscating HTTP traffic. It uses the HTTP Range method to receive resources in chunks, TCP
MSS to limit the size of individual chunks, and advertised window size to control the flow of chunks
in transmission.</li>
          <li>Panchenko et al. <xref target="panchenko2011website" format="default"/> developed Decoy, which is a simple mechanism that loads
a benign page alongside a real page. This seeks to mask the real page load by properties of the "decoy" page.
As with morphing, this defense requires application-specific knowledge about benign pages and is best
implemented outside of the transport layer.</li>
          <li>The Tor project implemented HTTP pipelining <xref target="perry2011experimental" format="default"/>, which bundles egress HTTP/1.1
requests into batches of varying sizes with random orders. Batching requests to mask request and response sizes
could be made easier with HTTP/2 <xref target="RFC7540" format="default"/>, HTTP/3, and QUIC, since these protocol naturally support
multiplexing. However, pipelining and batching may necessarily introduce latency delays that negatively impact
the user experience.</li>
          <li>Cherubin et al. <xref target="cherubin2017website" format="default"/> design two application-layer defenses called Application
Layer Padding Concerns Adversaries (ALPaCA) and Lightweight application-Layer Masquerading Add-on (LLaMA).
ALPaCA is a server-side defense that pads first-party content (deterministically or probabilistically)
according to a known distribution. (Deterministic padding similar to Tamaraw performs worse than
probabilistic padding.) LLaMA is similar to randomized pipelining, yet differs in that requests are also
delayed (if necessary) and spurious requests are generated according to some probability distribution.
Comparatively, ALPaCA yields a greater reduction in WF attack accuracy than LLaMA.</li>
          <li>Lu et al. <xref target="lu2018dynaflow" format="default"/> designed DynaFlow, which is a defense that dynamically adjusts
traffic flows using a combination of burst pattern morphing, constant traffic flow with flexible
intervals, and burst padding. DynaFlow overhead is 40% less than that of Tamaraw and was shown
to have similar benefits.</li>
          <li>Rahman <xref target="rahman18gan" format="default"/> uses generative adversarial networks (GANs) to modify candidate burst properties
of packet traces, i.e., by inserting dummy packets, such that they appear indistinguishable from other traces.
Normally, the generator component in a GAN uses random noise to produce information that matches a target
data distribution as classified by the discriminator component. Rahman uses a modified GAN architecture
wherein the generator produces padding (dummy packets) for input data such that the discriminator cannot
distinguish it from noise, or a desired burst packet sequence. Preliminary results with the GAN trained and
tested on defended traffic, i.e., traffic already subject to some form of WF defense, show a 9% increase in
bandwidth and 15% decrease in attack accuracy (from 98% to 85% in a closed world setting).</li>
          <li>Imani et al. <xref target="imanimockingbird" format="default"/> developed Mockingbird, a defense built on using generated adversarial
examples, i.e., dummy traffic designed to disrupt classifier behavior, aimed towards model misclassification.
When run on classifiers trained without adversarial examples, Mockingbird reduced state-of-the-art DF attacks
and CUMUL attacks from <xref target="panchenko2016website" format="default"/> from 98% to 3% and 92% to 31%, respectively. Conversely, classifiers
trained and hardened with adversarial examples only reduce attack accuracy from 95% to between 25-57%, respectively.
Classification results for half-duplex traces, i.e., those in which traffic flows in half-duplex mode, are lower.
Mockingbird's bandwidth overhead is tunable based on parameters that control the internal traffic shaping algorithm.</li>
          <li>Gong et al. developed <xref target="gong2020zero" format="default"/> is a lightweight defense that does not delay any packets, minimizing its effect on
user experience. Instead of adding packets during a packet trace in order to obfuscate which page it came from, GLUE adds packets between
packet traces (during user think time/downtime) to merge them together, creating a seamless sequence of packets covering
multiple page loads. Attackers are unable to train classifiers for multiple contiguous traces and also unable to identify
individual page traces from the sequence. This is in part because the GLUE used is itself a real packet trace, thwarting
attacker classification. GLUE also adds extra noise packets ("FRONT") in the first trace as it is vulnerable otherwise.</li>
        </ul>
      </section>
      <section anchor="traffic-splitting" numbered="true" toc="default">
        <name>Traffic Splitting</name>
        <t>Traffic splitting is a type of defense wherein application data is sent over multiple, disjoint network paths.
Multipath TCP (MPTCP) is one type of "traffic splitting" protocol, wherein an endpoint may send TCP segments for a
single connection over multiple interfaces. This is commonly done for multi-homed devices, such as mobile
devices with cellular and WiFi or wired network connections. Traffic splitting assumes that guided traffic distribution
reduces information available to an adversary, and thereby decreases the success probability of WF attacks. Traffic
splitting defenses differ in the algorithm used for traffic distribution.</t>
        <t>Henri et al. <xref target="henri2020protecting" format="default"/> studied several traffic splitting algorithms, including: weighted and non-weighted
round-robin path splitting, incoming and outgoing path split, fixed-probability splitting, and variants of per-connection
uniform probability splitting. The best results came from a per-connection path splitting variant where the
maximum number of packets sent on any given path was limited by a random variable chosen from a geometric distribution.
(Once this limit was reached, a new path was selected uniformly at random.) De la Cadena et al. <xref target="de2019poster" format="default"/>
also study path splitting algorithmss. They conclude that a weighted random path selection algorithm works best.
(The authors do not give specifics of path weight probability derivation.)</t>
      </section>
    </section>
    <section anchor="open-problems-and-directions" numbered="true" toc="default">
      <name>Open Problems and Directions</name>
      <t>To date, WF attacks target clients running over Tor or some other anonymizing service, meaning that WF
attacks are likely more accurate on normal TLS-protected connections. Moreover, attacks normally assume clients
use HTTP/1.1 with parallel connections for parallel resource fetches. In recent years, however, protocols
such as SPDY, HTTP/2, and QUIC with built-in padding support and multiplexed stream-based connections
should make existing attacks more difficult to carry out. That said, it is unclear how exactly these protocol
design trends will impact WF attacks. A non-exhaustive list of questions that warrant further research
are below:</t>
      <ol spacing="normal" type="1">
        <li>How does connection coalescing and consolidation affect WF attacks? Technologies such as DNS-over-HTTPS
and ESNI favor architectures wherein a single network or connection can serve multiple origins or resources.
With connection coalescing, traffic for multiple resources is sent on the same connection, thereby adding
effects similar to that of the Decoy defense mechanism described in <xref target="defenses" format="default"/></li>
        <li>To what extent does protocol multiplexing increase WF attack difficulty? Using a single connection
with multiple streams to avoid HoL blocking saves on connection startup and bandwidth costs while simultaneously
mixing information from multiple requests and resources on the same connection.</li>
        <li>How can protocol features such as HTTP Push be used to improve WF defense efficacy? Defenses without
cooperative peer support often induce suboptimal bandwidth or latency costs. If both endpoints of a connection
participate in the defense, even proactively with Push, perhaps this could be improved.</li>
        <li>Can connection bootstrapping techniques such as those used by ESNI be used to distribute WF defense
information? One possible approach is to distribute client padding profiles derived from CDN knowledge
of serviced resources.</li>
        <li>How can clients build, use, and possibly share WF defense information to benefit others?</li>
        <li>How can applications package websites and subresources in such a way that limits unique information?
For example, websites link to third party resources in an ad-hoc fashion, causing the subsequent trace of
browser fetches to possibly uniquely identify the website.</li>
      </ol>
      <t>Research into the above questions will help the IETF community better understand the extent to which
WF attacks are a problem for Internet users in general.</t>
      <t>It is worth mentioning that traffic-based WF attacks may not be required to achieve the desired
goal of learning a connection's destination. Network connections by nature reveal information
about endpoint behavior. The relationship between network address and domains, especially when
stable and unique, are a strong signal for website fingerprinting. Trevisan et al. <xref target="trevisan2016towards" format="default"/>
explored use of this signal as a reliable mechanism for website fingerprinting. They find that
most major services (domains) have clearly associated IP address(es), though
these addresses may change over time. Jiang et al. <xref target="jiang2007lightweight" format="default"/>
and Tammaro et al. <xref target="tammaro2012exploiting" format="default"/> also previously came to the same
conclusion. Address-based website fingerprinting was also explored by Patil and
Borisov <xref target="patil2019can" format="default"/>, wherein they showed that addresses, especially when
grouped together as part of a single web page load, leak a substantial amount of
information about the corresponding domain. Thus, in general, classifiers that rely
solely on network addresses may be used to aid website fingerprinting attacks.</t>
    </section>
    <section anchor="protocol-design-considerations" numbered="true" toc="default">
      <name>Protocol Design Considerations</name>
      <t>New protocols such as TLS 1.3 and QUIC are designed with privacy-protections in mind.
TLS 1.3, for example, supports record-layer padding <xref target="RFC8446" format="default"/>, although it is
not used widely in practice. Despite this, TLS connections still leak metadata, including
negotiatied ciphersuites. (See <xref target="fordTLSMetadata" format="default"/> for a discussion of this issue.)
QUIC is more aggressive in its use of encryption as both a mitigation for middlebox
ossificatiion and privacy enhancement. IPsec Traffic Flow Confidentiality <xref target="RFC4303" format="default"/>
and Traffic Flow Security <xref target="I-D.ietf-ipsecme-iptfs" format="default"/> are two mechanisms by which
endpoints can ESP datagrams to mask size metadata.</t>
      <t>Future protocols should follow these trends when possible to remove unnecessary metadata from
the network.</t>
    </section>
    <section anchor="security-considerations" numbered="true" toc="default">
      <name>Security Considerations</name>
      <t>This document surveys security and privacy attacks and defenses on encrypted TLS connections.
It does not introduce, specify, or recommend any particular mitigation to the aforementioned
attacks.</t>
    </section>
    <section anchor="iana-considerations" numbered="true" toc="default">
      <name>IANA Considerations</name>
      <t>This document makes no IANA requests.</t>
    </section>
  </middle>
  <back>
    <references>
      <name>Informative References</name>
      <reference anchor="gong2020zero" target="https://www.usenix.org/system/files/sec20summer_gong_prepub.pdf">
        <front>
          <title>Zero-delay Lightweight Defenses against Website Fingerprinting</title>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="CRIME" target="https://www.ekoparty.org/archive/2012/CRIME_ekoparty2012.pdf">
        <front>
          <title>The CRIME Attack</title>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="TIME">
        <front>
          <title>A Perfect CRIME? Only TIME Will Tell</title>
          <seriesInfo name="Black Hat Europe 2013" value=""/>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="trevisan2016towards">
        <front>
          <title>Towards web service classification using addresses and DNS</title>
          <seriesInfo name="Wireless Communications and Mobile Computing Conference (IWCMC), 2016 International. IEEE, 2016" value=""/>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="jiang2007lightweight">
        <front>
          <title>Lightweight application classification for network management</title>
          <seriesInfo name="SIGCOMM workshop on Internet network management, 2007" value=""/>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="tammaro2012exploiting">
        <front>
          <title>Exploiting packet-sampling measurements for traffic characterization and classification</title>
          <seriesInfo name="International Journal of Network Management, 2012" value=""/>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="foremski2014dns">
        <front>
          <title>DNS-Class -- immediate classification of IP flows using DNS</title>
          <seriesInfo name="International Journal of Network Management" value=""/>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="gong2010fingerprinting">
        <front>
          <title>Fingerprinting websites using remote traffic analysis</title>
          <seriesInfo name="Proceedings of the 17th ACM conference on Computer and communications security" value=""/>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="li2018can">
        <front>
          <title>Can We Learn What People Are Doing from Raw DNS Queries?</title>
          <seriesInfo name="IEEE INFOCOM 2018-IEEE Conference on Computer Communications" value=""/>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="nithyanand2014glove">
        <front>
          <title>Glove -- A bespoke website fingerprinting defense</title>
          <seriesInfo name="Proceedings of the 13th Workshop on Privacy in the Electronic Society" value=""/>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="pironti2012identifying">
        <front>
          <title>Identifying website users by TLS traffic analysis -- New attacks and effective countermeasures</title>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="backes2013preventing">
        <front>
          <title>Preventing Side-Channel Leaks in Web Traffic -- A Formal Approach</title>
          <seriesInfo name="NDSS, 2013" value=""/>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="coull2007web">
        <front>
          <title>On Web Browsing Privacy in Anonymized NetFlows</title>
          <seriesInfo name="USENIX Security Symposium" value=""/>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="hayes2016k">
        <front>
          <title>k-fingerprinting -- A Robust Scalable Website Fingerprinting Technique</title>
          <seriesInfo name="USENIX Security Symposium, 2016" value=""/>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="danezis2009traffic">
        <front>
          <title>Traffic Analysis of the HTTP Protocol over TLS</title>
          <seriesInfo name="2009" value=""/>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="wang2014comparing">
        <front>
          <title>Comparing website fingerprinting attacks and defenses</title>
          <seriesInfo name="Technical Report 2013-30, CACR, 2013." value=""/>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="juarez2014critical">
        <front>
          <title>A critical evaluation of website fingerprinting attacks</title>
          <seriesInfo name="ACM SIGSAC Conference on Computer and Communications Security, 2014" value=""/>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="cherubin2017website">
        <front>
          <title>Website fingerprinting defenses at the application layer</title>
          <seriesInfo name="Privacy Enhancing Technologies, 2017" value=""/>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="lu2018dynaflow">
        <front>
          <title>DynaFlow -- An Efficient Website Fingerprinting Defense Based on Dynamically-Adjusting Flows</title>
          <seriesInfo name="Workshop on Privacy in the Electronic Society, 2018" value=""/>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="shulman2014pretty">
        <front>
          <title>Pretty bad privacy -- Pitfalls of DNS encryption</title>
          <seriesInfo name="Workshop on Privacy in the Electronic Society, 2014" value=""/>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="siby2018dns">
        <front>
          <title>DNS Privacy not so private -- the traffic analysis perspective</title>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="rahman2018using">
        <front>
          <title>Using Packet Timing Information in Website Fingerprinting</title>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="sirinam2018deep">
        <front>
          <title>Deep fingerprinting -- Undermining website fingerprinting defenses with deep learning</title>
          <seriesInfo name="arXiv preprint arXiv:1801.02265" value=""/>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="wang2016website">
        <front>
          <title>Website fingerprinting -- Attacks and defenses</title>
          <seriesInfo name="University of Waterloo" value=""/>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="wagner1996analysis">
        <front>
          <title>Analysis of the SSL 3.0 protocol</title>
          <seriesInfo name="USENIX Workshop on Electronic Commerce Proceedings, 1996" value=""/>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="cheng1998traffic">
        <front>
          <title>Traffic analysis of SSL encrypted web browsing</title>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="sun2002statistical">
        <front>
          <title>Statistical identification of encrypted web browsing traffic</title>
          <seriesInfo name="IEEE, 2002" value=""/>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="hintz2002fingerprinting">
        <front>
          <title>Fingerprinting websites using traffic analysis</title>
          <seriesInfo name="International Workshop on Privacy Enhancing Technologies, 2002" value=""/>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="shmatikov2006timing">
        <front>
          <title>Timing analysis in low-latency mix networks -- Attacks and defenses</title>
          <seriesInfo name="European Symposium on Research in Computer Security, 2006" value=""/>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="bissias2005privacy">
        <front>
          <title>Privacy vulnerabilities in encrypted HTTP streams</title>
          <seriesInfo name="International Workshop on Privacy Enhancing Technologies, 2005" value=""/>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="liberatore2006inferring">
        <front>
          <title>Inferring the source of encrypted HTTP connections</title>
          <seriesInfo name="ACM Conference on Computer and Communications Security, 2006" value=""/>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="herrmann2009website">
        <front>
          <title>Website fingerprinting -- attacking popular privacy enhancing technologies with the multinomial naive-bayes classifier</title>
          <seriesInfo name="ACM workshop on Cloud computing security, 2009" value=""/>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="lu2010website">
        <front>
          <title>Website fingerprinting and identification using ordered feature sequences</title>
          <seriesInfo name="European Symposium on Research in Computer Security, 2010" value=""/>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="panchenko2011website">
        <front>
          <title>Website fingerprinting in onion routing based anonymization networks</title>
          <seriesInfo name="ACM workshop on Privacy in the electronic society, 2011" value=""/>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="panchenko2016website" target="https://www.freehaven.net/anonbib/cache/fingerprinting-ndss2016.pdf">
        <front>
          <title>Website Fingerprinting at Internet Scale</title>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="cai2012touching">
        <front>
          <title>Touching from a distance -- Website fingerprinting attacks and defenses</title>
          <seriesInfo name="ACM conference on Computer and communications security, 2012" value=""/>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="wang2013improved">
        <front>
          <title>Improved website fingerprinting on tor</title>
          <seriesInfo name="Workshop on privacy in the electronic society, 2013" value=""/>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="wang2014effective">
        <front>
          <title>Effective Attacks and Provable Defenses for Website Fingerprinting</title>
          <seriesInfo name="USENIX Security Symposium, 2014" value=""/>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="abe2016fingerprinting">
        <front>
          <title>Fingerprinting attack on tor anonymity using deep learning</title>
          <seriesInfo name="Asia-Pacific Advanced Network, 2016" value=""/>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="rimmer2018automated">
        <front>
          <title>Automated website fingerprinting through deep learning</title>
          <seriesInfo name="Network &amp; Distributed System Security Symposium (NDSS), 2018" value=""/>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="yan2018feature">
        <front>
          <title>Feature selection for website fingerprinting</title>
          <seriesInfo name="Proceedings on Privacy Enhancing Technologies, 2018" value=""/>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="bhat2018var">
        <front>
          <title>Var-CNN and DynaFlow -- Improved Attacks and Defenses for Website Fingerprinting</title>
          <seriesInfo name="arXiv preprint arXiv:1802.10215" value=""/>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="oh2017pfp">
        <front>
          <title>p-FP -- Extraction, Classification, and Prediction of Website Fingerprints with Deep Learning</title>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="reed2017identifying">
        <front>
          <title>Identifying https-protected netflix videos in real-time</title>
          <seriesInfo name="ACM on Conference on Data and Application Security and Privacy, 2017" value=""/>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="schuster2017beauty">
        <front>
          <title>Beauty and the burst -- Remote identification of encrypted video streams</title>
          <seriesInfo name="USENIX Security, 2017" value=""/>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="miller2014know">
        <front>
          <title>I know why you went to the clinic -- Risks and realization of https traffic analysis</title>
          <seriesInfo name="International Symposium on Privacy Enhancing Technologies Symposium, 2014" value=""/>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="wright2009traffic">
        <front>
          <title>Traffic Morphing -- An Efficient Defense Against Statistical Traffic Analysis</title>
          <seriesInfo name="NDSS, 2009" value=""/>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="luo2011httpos">
        <front>
          <title>HTTPOS -- Sealing Information Leaks with Browser-side Obfuscation of Encrypted Flows</title>
          <seriesInfo name="NDSS, 2011" value=""/>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="perry2011experimental" target="https://blog.torproject.org/experimental-defense-website-traffic-fingerprinting">
        <front>
          <title>Experimental defense for website traffic fingerprinting</title>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="dyer2012peek">
        <front>
          <title>Peek-a-boo, i still see you -- Why efficient traffic analysis countermeasures fail</title>
          <seriesInfo name="IEEE Symposium on Security and Privacy, 2012" value=""/>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="wang2015walkie">
        <front>
          <title>Walkie-talkie -- An effective and efficient defense against website fingerprinting</title>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="juarez2015wtf" target="https://pdfs.semanticscholar.org/0f54/4d0845cb9f317722759dc49e1493ef30d83d.pdf">
        <front>
          <title>WTF-PAD -- toward an efficient website fingerprinting defense for tor</title>
          <seriesInfo name="CoRR, abs/1512.00524" value=""/>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="juarez2016toward">
        <front>
          <title>Toward an efficient website fingerprinting defense</title>
          <seriesInfo name="European Symposium on Research in Computer Security, 2016" value=""/>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="fordTLSMetadata" target="http://bford.info/pub/net/tlsmeta.pdf">
        <front>
          <title>Metadata Protection Considerations for TLS Present and Future</title>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="rahman2019tik" target="https://arxiv.org/pdf/1902.06421.pdf">
        <front>
          <title>Tik-Tok -- The Utility of Packet Timing in Website Fingerprinting Attacks</title>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="rahman18gan" target="https://www.rahmanmsaidur.com/projects/Fall_18_Generating_Adversarial_Packets.pdf">
        <front>
          <title>Generating Adversarial Packets for Website Fingerprinting Defense</title>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="imanimockingbird" target="https://arxiv.org/pdf/1902.06626.pdf">
        <front>
          <title>Mockingbird -- Defending Against Deep-Learning-Based Website Fingerprinting Attacks with Adversarial Traces</title>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="patil2019can" target="https://irtf.org/anrw/2019/anrw2019-final44-acmpaginated.pdf">
        <front>
          <title>What can you learn from an IP?</title>
          <author>
            <organization/>
          </author>
          <date>n.d.</date>
        </front>
      </reference>
      <reference anchor="de2019poster" target="https://dl.acm.org/doi/10.1145/3319535.3363249">
        <front>
          <title>Traffic Splitting to Counter Website Fingerprinting</title>
          <author>
            <organization/>
          </author>
          <date year="2019"/>
        </front>
      </reference>
      <reference anchor="henri2020protecting" target="https://petsymposium.org/2020/files/papers/issue2/popets-2020-0019.pdf">
        <front>
          <title>Protecting against Website Fingerprinting with Multihoming</title>
          <author>
            <organization/>
          </author>
          <date year="2020"/>
        </front>
      </reference>
      <reference anchor="RFC8446" target="https://www.rfc-editor.org/info/rfc8446">
        <front>
          <title>The Transport Layer Security (TLS) Protocol Version 1.3</title>
          <seriesInfo name="DOI" value="10.17487/RFC8446"/>
          <seriesInfo name="RFC" value="8446"/>
          <author initials="E." surname="Rescorla" fullname="E. Rescorla">
            <organization/>
          </author>
          <date year="2018" month="August"/>
          <abstract>
            <t>This document specifies version 1.3 of the Transport Layer Security (TLS) protocol.  TLS allows client/server applications to communicate over the Internet in a way that is designed to prevent eavesdropping, tampering, and message forgery.</t>
            <t>This document updates RFCs 5705 and 6066, and obsoletes RFCs 5077, 5246, and 6961.  This document also specifies new requirements for TLS 1.2 implementations.</t>
          </abstract>
        </front>
      </reference>
      <reference anchor="I-D.ietf-quic-transport" target="http://www.ietf.org/internet-drafts/draft-ietf-quic-transport-29.txt">
        <front>
          <title>QUIC: A UDP-Based Multiplexed and Secure Transport</title>
          <seriesInfo name="Internet-Draft" value="draft-ietf-quic-transport-29"/>
          <author initials="J" surname="Iyengar" fullname="Jana Iyengar">
            <organization/>
          </author>
          <author initials="M" surname="Thomson" fullname="Martin Thomson">
            <organization/>
          </author>
          <date month="June" day="9" year="2020"/>
          <abstract>
            <t>This document defines the core of the QUIC transport protocol. Accompanying documents describe QUIC's loss detection and congestion control and the use of TLS for key negotiation.  Note to Readers  Discussion of this draft takes place on the QUIC working group mailing list (quic@ietf.org (mailto:quic@ietf.org)), which is archived at https://mailarchive.ietf.org/arch/search/?email_list=quic  Working Group information can be found at https://github.com/quicwg; source code and issues list for this draft can be found at https://github.com/quicwg/base-drafts/labels/-transport.</t>
          </abstract>
        </front>
      </reference>
      <reference anchor="RFC8484" target="https://www.rfc-editor.org/info/rfc8484">
        <front>
          <title>DNS Queries over HTTPS (DoH)</title>
          <seriesInfo name="DOI" value="10.17487/RFC8484"/>
          <seriesInfo name="RFC" value="8484"/>
          <author initials="P." surname="Hoffman" fullname="P. Hoffman">
            <organization/>
          </author>
          <author initials="P." surname="McManus" fullname="P. McManus">
            <organization/>
          </author>
          <date year="2018" month="October"/>
          <abstract>
            <t>This document defines a protocol for sending DNS queries and getting DNS responses over HTTPS.  Each DNS query-response pair is mapped into an HTTP exchange.</t>
          </abstract>
        </front>
      </reference>
      <reference anchor="I-D.ietf-tls-esni" target="http://www.ietf.org/internet-drafts/draft-ietf-tls-esni-07.txt">
        <front>
          <title>TLS Encrypted Client Hello</title>
          <seriesInfo name="Internet-Draft" value="draft-ietf-tls-esni-07"/>
          <author initials="E" surname="Rescorla" fullname="Eric Rescorla">
            <organization/>
          </author>
          <author initials="K" surname="Oku" fullname="Kazuho Oku">
            <organization/>
          </author>
          <author initials="N" surname="Sullivan" fullname="Nick Sullivan">
            <organization/>
          </author>
          <author initials="C" surname="Wood" fullname="Christopher Wood">
            <organization/>
          </author>
          <date month="June" day="1" year="2020"/>
          <abstract>
            <t>This document describes a mechanism in Transport Layer Security (TLS) for encrypting a ClientHello message under a server public key.</t>
          </abstract>
        </front>
      </reference>
      <reference anchor="RFC7540" target="https://www.rfc-editor.org/info/rfc7540">
        <front>
          <title>Hypertext Transfer Protocol Version 2 (HTTP/2)</title>
          <seriesInfo name="DOI" value="10.17487/RFC7540"/>
          <seriesInfo name="RFC" value="7540"/>
          <author initials="M." surname="Belshe" fullname="M. Belshe">
            <organization/>
          </author>
          <author initials="R." surname="Peon" fullname="R. Peon">
            <organization/>
          </author>
          <author initials="M." surname="Thomson" fullname="M. Thomson" role="editor">
            <organization/>
          </author>
          <date year="2015" month="May"/>
          <abstract>
            <t>This specification describes an optimized expression of the semantics of the Hypertext Transfer Protocol (HTTP), referred to as HTTP version 2 (HTTP/2).  HTTP/2 enables a more efficient use of network resources and a reduced perception of latency by introducing header field compression and allowing multiple concurrent exchanges on the same connection.  It also introduces unsolicited push of representations from servers to clients.</t>
            <t>This specification is an alternative to, but does not obsolete, the HTTP/1.1 message syntax.  HTTP's existing semantics remain unchanged.</t>
          </abstract>
        </front>
      </reference>
      <reference anchor="RFC4303" target="https://www.rfc-editor.org/info/rfc4303">
        <front>
          <title>IP Encapsulating Security Payload (ESP)</title>
          <seriesInfo name="DOI" value="10.17487/RFC4303"/>
          <seriesInfo name="RFC" value="4303"/>
          <author initials="S." surname="Kent" fullname="S. Kent">
            <organization/>
          </author>
          <date year="2005" month="December"/>
          <abstract>
            <t>This document describes an updated version of the Encapsulating Security Payload (ESP) protocol, which is designed to provide a mix of security services in IPv4 and IPv6.  ESP is used to provide confidentiality, data origin authentication, connectionless integrity, an anti-replay service (a form of partial sequence integrity), and limited traffic flow confidentiality.  This document obsoletes RFC 2406 (November 1998).  [STANDARDS-TRACK]</t>
          </abstract>
        </front>
      </reference>
      <reference anchor="I-D.ietf-ipsecme-iptfs" target="http://www.ietf.org/internet-drafts/draft-ietf-ipsecme-iptfs-01.txt">
        <front>
          <title>IP Traffic Flow Security</title>
          <seriesInfo name="Internet-Draft" value="draft-ietf-ipsecme-iptfs-01"/>
          <author initials="C" surname="Hopps" fullname="Christian Hopps">
            <organization/>
          </author>
          <date month="March" day="2" year="2020"/>
          <abstract>
            <t>This document describes a mechanism to enhance IPsec traffic flow security by adding traffic flow confidentiality to encrypted IP encapsulated traffic.  Traffic flow confidentiality is provided by obscuring the size and frequency of IP traffic using a fixed-sized, constant-send-rate IPsec tunnel.  The solution allows for congestion control as well.</t>
          </abstract>
        </front>
      </reference>
    </references>
    <section anchor="acknowledgements" numbered="true" toc="default">
      <name>Acknowledgements</name>
      <t>The authors would like to thank Frederic Jacobs and Tim Taubert for feedback
on earlier versions of this document.</t>
    </section>
  </back>
  <!-- ##markdown-source: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-->

</rfc>
