PreprintPDF Available

Fake News Detection via NLP is Vulnerable to Adversarial Attacks


Abstract and Figures

News plays a significant role in shaping people's beliefs and opinions. Fake news has always been a problem, which wasn't exposed to the mass public until the past election cycle for the 45th President of the United States. While quite a few detection methods have been proposed to combat fake news since 2015, they focus mainly on linguistic aspects of an article without any fact checking. In this paper, we argue that these models have the potential to misclassify fact-tampering fake news as well as under-written real news. Through experiments on Fakebox, a state-of-the-art fake news detector, we show that fact tampering attacks can be effective. To address these weaknesses, we argue that fact checking should be adopted in conjunction with linguistic characteristics analysis, so as to truly separate fake news from real news. A crowdsourced knowledge graph is proposed as a straw man solution to collecting timely facts about news events.
Content may be subject to copyright.
Fake News Detection via NLP is Vulnerable to Adversarial Attacks
Zhixuan Zhou1,2, Huankang Guan1, Meghana Moorthy Bhat2and Justin Hsu2
1Hongyi Honor College, Wuhan University, Wuhan, China
2Department of Computer Science, University of Wisconsin-Madison, Madison, USA
{kyriezoe, hkguan}, {mbhat2, justhsu}
Keywords: Fake News Detection, NLP, Attack, Fact Checking, Outsourced Knowledge Graph
Abstract: News plays a significant role in shaping people’s beliefs and opinions. Fake news has always been a problem,
which wasn’t exposed to the mass public until the past election cycle for the 45th President of the United States.
While quite a few detection methods have been proposed to combat fake news since 2015, they focus mainly
on linguistic aspects of an article without any fact checking. In this paper, we argue that these models have the
potential to misclassify fact-tampering fake news as well as under-written real news. Through experiments on
Fakebox, a state-of-the-art fake news detector, we show that fact tampering attacks can be effective. To address
these weaknesses, we argue that fact checking should be adopted in conjunction with linguistic characteristics
analysis, so as to truly separate fake news from real news. A crowdsourced knowledge graph is proposed as a
straw man solution to collecting timely facts about news events.
Fake news is an increasingly common feature of to-
day’s political landscape. To help address this issue,
researchers and media experts have proposed fake
news detectors adopting natural language processing
(NLP) to analyze word patterns and statistical corre-
lations of news articles. While these detectors achieve
impressive accuracy on existing examples of manip-
ulated news, the analysis is typically quite shallow—
roughly, models check whether news articles conform
to standard norms and styles used by professional
journalists. This leads to two drawbacks.
First, these models can detect fake news only
when they are under-written, for instance when the
content is totally unrelated to the headline (so-called
“clickbait”) or when the article includes words con-
sidered to be biased or inflammatory. While this
criteria suffices to detect many existing examples of
fake news, more sophisticated rumor disseminators
can craft more subtle attacks, for instance taking a
well-written real news article and tampering the ar-
ticle in a targeted way. By preserving the original
subject matter and relating the content tightly to the
headline without using biased phrases, an adversar-
ial article can easily evade detection. To demon-
strate this kind of attack, we evaluate a state-of-the-art
model called Fakebox. We introduce three classes of
attacks: fact distortion, subject-object exchange and
cause confounding. We generate adversarial versions
of real news from a dataset by McIntire (2018), and
show that Fakebox achieves low accuracy when clas-
sifying these examples.
At the same time, requirements posed by current
detectors are often too strict. Real news which is
under-written or talks about certain political and re-
ligious topics is likely to be mistakenly rejected, re-
gardless of its accuracy. This is a particularly seri-
ous problem for open platforms, such as Twitter in
the United States and TouTiao in China, where much
of the news is contributed by users with diverse back-
grounds. To prevent frustrating false positives, plat-
forms are still heavily relying on manual work for
separating fake news from real news. We provide ex-
perimental evidence for Fakebox’s potential of mis-
classifying real news.
Taken together, our experiments highlight vulner-
able aspects of fake news detection methods based
purely on NLP. Without deeper semantic knowledge,
such detectors are easily fooled by fact-tampering at-
tacks and can suffer from a high rate of false pos-
itives, mistakenly classifying under-written yet real
news which may not be written in a journalistic style.
To address these problems, we argue that some form
of fact-based knowledge must be adopted alongside
NLP-based models. What this knowledge is remains
to be seen, but we consider a straw man solution: a
crowdsourced knowledge graph that aggregates infor-
Table 1: Examples of fact tampering attacks.
Attack type Original Adversarial
Fact distortion 12 people were injured in the shooting. 24 people were killed in the shooting.
Subject-object exchange A gangster was shot by the police. A policeman was shot by the gangster.
Cause confounding
The condom policy originated in
1992 . . . The Boy Scouts have de-
cided to accept people who identify
as gay and lesbian. (unrelated events)
The inclusion of gays, lesbians and
girls in the Boy Scouts led to the con-
dom policy.
mation about news events and helps judge whether
information extracted from news articles is reliable.
The rest of our paper is organized as follows. In
Section 2, we introduce three kinds of attacks tar-
geting fake news detectors. Section 3 introduces the
dataset from which we generate adversarial examples
and our target model Fakebox. We evaluate Fakebox’s
performance under attacks in Section 4. In Section 5,
we discuss weaknesses of NLP-based detectors and
propose augmenting detectors with knowledge infor-
mation, such as a crowdsourced knowledge graph, as
a step towards making fake news detection more ro-
bust. Finally, we discuss related work in Section 6
and conclude in Section 7.
Adversarial Machine Learning is an emerging field
of applied machine learning that seeks to understand
how machine learning classifiers can be attacked by
malicious users. To see how well existing fake news
detectors perform against adversarial inputs, we ex-
plored three kinds of adversarial examples with tam-
pering focusing on different aspects of an article:
Fact distortion: exaggerating or modifying on
some words. Character, time, location, relation,
extent and any other element can be distorted;
Subject-object exchange: with this attack readers
will be confused as to who is the performer and
who is the receiver of an action. It can be per-
formed on sentence level;
Cause confounding: either building non-existent
causal relationship between two independent
events, or cutting off some parts of a story, leaving
only the parts that an adversarial wants to present
to his readers.
Examples of these attacks are shown in Table 1. By
repeating these modifications, we can significantly
change the semantic content of a news article without
distorting the “writing style” of the original one—the
modified article is still presented in a seemingly logi-
cal and sound way.
3.1 Dataset
We generate adversarial examples from articles in
McIntire’s fake-real-news-dataset, an open-source
dataset extensively used in misinformation research.
The dataset contains 6,335 articles. 3,171 of them are
labeled as real and 3,164 of them are labeled as fake.
The ratio of real and fake news articles is roughly 1:1.
Titles, contents and veracity labels are provided. The
dataset does not include URLs, but we are primarily
concerned with the textual content rather than exter-
nal links, which can be manipulated in many other
ways. We manually check veracity of the news by
comparing them with reputable sources to increase
our confidence that the labels are reasonable.
3.2 Fakebox
Fakebox analyzes linguistic characteristics of news
articles to assess whether they are likely to be real
news or not. By looking at different aspects of an ar-
ticle (title, content and URL), using NLP models and
training on a manually curated database, Fakebox can
successfully identify fake news. Edell (2018) reports
achieving classification accuracy upwards of 95%.
Fakebox checks several aspects of each article:
Title or headline: checked for clickbait;
Content: analyzed to determine whether it’s writ-
ten like real news;
Domain: some websites are known for hosting
certain types of content, like hoaxes and satires.
If an article is written like a real one, Fakebox labels it
as impartial and gives it a score between 60 and 100.
If an article is not written like a real one, Fakebox
labels it as biased and gives it a score between 0 and
Table 2: Normal-time output of Fakebox.
Labels Impartial Biased Unsure
Real news 1159 1477 535
Fake news 537 2184 443
Table 3: Normal-time accuracy of Fakebox with unsure cases excluded.
News type Number of articles Correctly classified Classification accuracy
Real 2636 1159 43.97%
Fake 2721 2184 80.26%
Total 5357 3343 62.40%
40. Otherwise, Fakebox labels it as unsure and gives
it a score between 40 and 60. It labels and assigns
quantitative scores for titles, contents and domains,
respectively. Higher-scoring articles are likely to be
more reliable.
The main focus of Fakebox is on linguistic charac-
teristics of a news article without any fact checking,
which potentially makes it vulnerable when facing
news which is written in a similar style to real news
but is not factual. To test this hypothesis, we per-
formed an experimental evaluation of Fakebox. We
establish a baseline by testing Fakebox with unmodi-
fied examples from McIntire’s dataset. Then, we ap-
ply our attacks described in Section 2. All experi-
ments are conducted on an Intel machine equipped
with quad-core 1.80 GHz CPU, 8GB RAM, 256GB
SSD and running Windows 10.
4.1 Baseline Performance
We first test baseline performance of Fakebox with
McIntire’s dataset. We feed 6,335 headlines and arti-
cles into Fakebox and get back corresponding labels.
We take special care of its output labels for content
veracity. Output of Fakebox is shown in Table 2.
Real and Fake are actual attributes of news articles
while Impartial, Biased and Unsure are labels given
by Fakebox. We use true positive (TP) to denote cor-
rectly classified fake news, true negative (TN) to de-
note correctly classified real news, false positive (FP)
to denote misclassified real news and false negative
(FN) to denote misclassified fake news. False positive
rate (FPR) and false negative rate (FNR) are defined
as below:
FPR =FP/(F P +T N)
Fakebox’s accuracy on McIntire’s dataset is
52.77%, false rate is 31.79% and for the other 15.44%
samples, Fakebox is unsure about their veracity.
While it is acceptable for fake news detectors to be
unsure for some articles and leave the hard tasks to
field experts, which is what happens in news plat-
forms moderation nowadays, its accuracy in our ex-
periment is still unsatisfactory even if we don’t take
unsure labels into consideration. It performs well
when dealing with fake news where false negative rate
is only 19.74%. But on the other hand, it labels more
real news as biased than as impartial. Quantitatively,
its false positive rate is 56.03%. Its overall accuracy
when excluding unsure cases is 62.40%. The result is
shown in Table 3.
We observe that in false positive cases, words that
tend to be regarded as “fake” include “anti”, “prison”,
“terror”, “Islamism” and “Trump”. This focus on sen-
sitive terms leads to a crude analysis. Though much
fake news emerges around these topics, it is not ap-
propriate to give a large “bias weight” to these words,
which is implemented in many state-of-the-art mod-
els. After all, there are equally many real news arti-
cles talking about these issues.
We also observe that many of the real articles mis-
clasified as fake can be regarded as under-written, i.e.,
they are not written in a journalistic style. They are
likely to be written by grass-root writers from Twitter
who are interested in social issues and are willing to
share their opinions instead of professional journal-
4.2 Attack-time Performance
We generate adversarial examples by hand from real
news that are also labeled as impartial by Fakebox.
This process can be seen as an adversary selecting real
news from authority sources like New York Times and
performing considerable tampering by himself.
For fact distortion, we simply substitute people,
places or actions without much effort. For exam-
ple, for the article titled “Is the GOP losing Wal-
mart?”, we substitute each “Walmart” in the content
with Apple”. The veracity score given by Fake-
box drops down by only 0.0073, which is negligi-
ble for its judgement. Tampering to other articles
as well doesn’t cause the veracity score to drop by
much. This kind of tiny tampering can have an out-
sized impact—imagine that company A is involved in
an information breach scandal but company B is re-
ported to be responsible by fake news.
For subject-object exchange tampering, the verac-
ity score doesn’t change at all, since term frequency
stays the same. This can be quite misleading: “a
gangster was shot by the police” and “a policeman
was shot by the gangster” are totally different and the
latter will cause public panic.
Cause checking is probably the most vulnerable
part of NLP-based detectors. For instance, there are
two real and unrelated articles labeled as impartial
by Fakebox, one about Walmart scoring 0.7151 for
veracity and the other about local politics in Cleve-
land scoring 0.7652. When we simply mix the two
articles together, the generated article is still labeled
as impartial and even reaches a much higher score
(0.8585). We further try to mix an article labeled as
impartial with an article labeled as biased and the ve-
racity score for the generated article is between the
scores of two original articles, which indicates that
only linguistic characteristics are inspected and facts
are never checked. As long as an adversary keeps ar-
ticles in a classical manner, he can mix totally unre-
lated events together, build non-existent causal rela-
tionships and evade detection.
As we can see in experiments, simply looking into
linguistic aspects is not enough for fake news detec-
tion. Two main defects of this method are its vulner-
ability to fact tampering attacks and its bias towards
under-written articles and certain topics. Given that
one of the essences of fake news is fact tampering,
fact checking could be quite helpful. However, it is
largely missing in current fake news detection mod-
5.1 A Straw Man Solution
There is urgent need to compare information ex-
tracted from news articles with “fact”, and source of
the fact is another key issue. Media and specialists
have delicate skills yet limited time and energy to col-
lect various fact from all sources. Fake news usually
comes on early stage after events happen and thus re-
quires early detecting, worsening the situation. One
possible solution to stopping fake news is to extract
key information from articles including causal rela-
tionships and compare it with a dynamically-updated
news fact knowledge graph. Such an approach was
also proposed by Pan et al. (2018).
A knowledge graph is a graph with entities of dif-
ferent types as nodes and various relations among
them as edges (Jia et al., 2016). Typical examples
include WordNet (Miller, 1995) and OpenKN (Jia
et al., 2014) and realistic applications include docu-
ment understanding (Wu et al., 2012) and link predic-
tion (Liu et al., 2014). Knowledge Graph is also used
by Google to enhance its search engine’s results with
information gathered from a variety of sources. The
information is presented to users in an infobox next
to the search results. An example Google knowledge
graph is shown in Figure 1.
Figure 1: Example knowledge graph.
Crowdsourcing is a distributed problem-solving
model in which a crowd of undefined size is en-
gaged to solve a complex problem through open calls
(Chatzimilioudis et al., 2012). It divides work be-
tween participants to achieve a cumulative result. It
is possible that a large crowd of non-experts can
collaborate well on a task that otherwise would re-
quire extensive efforts of a small group of experts
(Howe, 2006). The news aggregation site
(Mieghem, 2011) is another example of this method’s
application other than Wikipedia. While crowdsourc-
ing tends to result in high disagreement among con-
tributors, Dumitrache et al. (2018) showed that dis-
agreement is not noise but signal, and that in fact
crowdsourcing can not only be cheaper and scalable,
it can be higher quality and more informative as well,
for disagreement representation can be used to detect
low quality workers.
A crowdsourced generation of knowledge graphs
may be efficient and timely in the context of news
propagation. While fake news usually floods on the
early stage after an event happens, local or well-
informed people hear about the events faster and more
accurately. They can either be journalists or by-
standers who are equally responsible for fact main-
taining and fake news combating. If we can create a
structured visualized interface for building and edit-
ing knowledge graphs, where users only need to fill
in the “subject”, “action”, “object”, “time” and “loca-
tion” entities, they can easily fill in facts they are sure
of without much professional expertise. The design
could be visually similar to the Google knowledge
graph shown in Figure 1—which is friendly to non-
expert users—but it could work in a crowdsourced
manner. As the knowledge graph is updated dynami-
cally, timely fact information can be utilized to detect
fact tampering attacks in news articles.
The main drawback of our straw man solution is
the difficulty of collecting high-quality information.
While the crowdsourcing way of updating the knowl-
edge graph does ensure high efficiency, attackers with
special intentions have equal access to creating and
editing. If the fact entries collected with the knowl-
edge graph are not fact but “accomplices” created by
attackers, they cannot be utilized to help detect fake
news. How to address this issue is a serious challenge
when it comes to crowdsourcing.
5.2 What is a Fact?
It is highly difficult to give a clear and flexible def-
inition of “fact requirement” for different informa-
tion propagation contexts. In this work, we see “fact”
from a conventional, qualitative perspective and un-
derstand it as a statement that generally conforms to
a certain event or a piece of knowledge. Wikipedia
defines a fact as something that is consistent with ob-
jective reality or that can be proven with evidence—if
a statement can be demonstrated to correspond to ex-
perience, it’s a fact. However, we recognize that this
is by no means the end of the story and significantly
more research is needed to make the idea of a fact
more concrete.
Fake news detection became a hot research field in
2015. Since then, a number of methods have been put
forward. We list early, foundational works including
categorization of tasks and methods as well as build-
ing of platforms and datasets. We also list fake new
detection methods in three main categories.
6.1 Foundational Works
Rubin et al. (2015) separated the task of fake news
detection by type of fake: serious fabrications, large-
scale hoaxes and humorous fakes.
Conroy et al. (2015) provided a typology of ve-
racity assessment methods emerging from two major
categories—linguistic cue approaches and network
analysis approaches. They saw promise in an innova-
tive hybrid approach that combined the two methods.
Shao et al. (2016)) introduced Hoaxy, a platform
for the collection, detection, and analysis of online
misinformation, which had the potential to help peo-
ple understand the dynamics of real and fake news
sharing. Wang (2017) presented LIAR, a publicly
available dataset for fake news detection.
6.2 Linguistic Approaches
These models look simply at linguistic characteristics
such as grammar feature, word pattern, term count
and appearance of certain expressions. Defects of the
models are discussed in detail in our paper.
Chen et al. (2015) examined potential methods
for the automatic detection of clickbait. Methods for
recognizing both textual clickbaiting cues and non-
textual ones including image and user behavior were
Bourgonje et al. (2017) presented a system for de-
tecting the stance of headlines with regard to their cor-
responding article bodies. The approach could be ap-
plied in fake news, especially clickbait detection sce-
Granik and Mesyura (2017) showed a simple ap-
proach for fake news detection using naive Bayes
classifier and achieved decent result considering the
relative simplicity of their model.
Horne and Adali (2017) analyzed difference of
fake and real news in title features, complexity and
style of content. Elaboration Likelihood Model was
considered as a theory to explain the spread and per-
suasion of fake news.
Rashkin et al. (2017) compared language of real
news with that of satire, hoaxes, and propaganda to
find linguistic characteristics of untrustworthy text.
Their experiments adopted stylistic cues to help de-
termine the truthfulness of text.
6.3 Network Approaches
Models based on network analysis realize the impor-
tance of taking various background information into
account, instead of inspecting solely the articles them-
selves. They perform generally well at most times,
but when related information is missing or little, their
performance will drop.
Jin et al. (2016) improved news verification by
mining conflicting viewpoints in microblogs.
Long et al. (2017) proved that speaker profiles
such as party affiliation, speaker title, location and
credit history provided valuable information to vali-
date the credibility of news articles.
Farajtabar et al. (2017) proposed a multi-stage in-
tervention framework that tackled fake news in social
networks by combining reinforcement learning with a
point process network activity model.
Volkova et al. (2017) found social interaction fea-
tures were more informative for finer-grained sepa-
ration between four types of suspicious news (satire,
hoaxes, clickbait and propaganda) compared to syn-
tax and grammar features.
Tacchini et al. (2017) classified Facebook posts as
hoaxes or non-hoaxes with high accuracy on the basis
of the users who liked them.
6.4 Hybrid Approaches
Hybrid approaches combine the advantage of linguis-
tic models and network models, which intuitively out-
perform either of them.
Ruchansky et al. (2017) proposed a model that
combined the text of an article, the user response it
receives, and the source users promoting it for a more
accurate and automated prediction.
As far as we know, no other hybrid approaches are
available and fact-checking is absent from all existing
models. We also survey on commercial fake news de-
tectors and find that the majority of them take only
linguistic features into consideration.
In this paper, we evaluate a fake news detector Fake-
box on adversarial attacks, including fact-distortion,
subject-object exchange and cause confounding at-
tacks. Experiments show that our attack subverts the
model significantly. We believe that similar models
based solely on linguistic characteristics will perform
much less effectively in the real world and are espe-
cially vulnerable to tampering attacks. This kind of
attack is much more subtle, since it doesn’t change
the overall writing style of news articles and thus has
the potential to evade similarity detection. We argue
that multi-source fact comparing and checking must
be integrated into fake news detection models to truly
detect misinformation.
At the same time we find false positive rate rises
when it comes to either under-written real articles
or certain topics around which there is supposed to
be more fake news. The potential of misclassifying
under-written yet real news will hurt amateur news
writers’ enthusiasm. Thus we further suggest using
fact-checking as a helpful supplement so as to smooth
the negative effect of false positive judges.
One possible way to collect fact about news events
is to use a crowdsourced knowledge graph, which is
dynamically updated by local and well-informed peo-
ple. The timely information collected can then be
used to compare to that extracted from news articles
and help generate a label of veracity.
Our future work includes building a visualized in-
terface for news knowledge graph crowdsourcing, so
as to make work as easy as possible for non-experts
and stop fact-tampering fake news on early stage. We
also want to look at the issue of fake news propaga-
tion from a different angle, i.e., putting it in a social
context and examining human factors in order to bet-
ter understand the problem.
Bourgonje, P., Schneider, J. M., and Rehm, G. (2017).
From clickbait to fake news detection: An ap-
proach based on detecting the stance of headlines
to articles. In proceedings of the 2017 EMNLP.
Chatzimilioudis, G., Konstantinidis, A., Laoudias, C.,
and ZeinalipourYazti, D. (2012). Crowdsourcing
with smartphones. In IEEE Internet Comp., 36-
Chen, Y., Conroy, N. J., and Rubin, V. L. (2015). Mis-
leading online content: Recognizing clickbait as
false news. In Proceedings of the 2017 EMNLP.
Conroy, N. J., Rubin, V. L., and Chen, Y. (2015). Au-
tomatic deception detection: methods for finding
fake news. In Proceedings of the 78th ASIS&T
Annual Meeting.
Dumitrache, A., Aroyo, L., and Welty, C. (2018).
Crowdsourcing ground truth for medical relation
extraction. In ACM Transactions on Interactive
Intelligent Systems 8, 2.
Edell, A. (2018). I trained fake news detection ai with
>95% accuracy, and almost went crazy.
Farajtabar, M., Yang, J., Ye, X., Xu, H., Trivedi, R.,
Khalil, E., Li, S., Song, L., and Zha, H. (2017).
Fake news mitigation via point process based in-
tervention. In arXiv:1703.07823 [cs.LG].
Granik, M. and Mesyura, V. (2017). Fake news de-
tection using naive bayes classifier. In IEEE First
Ukraine Conference on Electrical and Computer
Engineering (UKRCON).
Horne, B. D. and Adali, S. (2017). This just in: Fake
news packs a lot in title, uses simpler, repeti-
tive content in text body, more similar to satire
than real news. In 2nd International Workshop
on News and Public Opinion at ICWSM.
Howe, J. (2006). The rise of crowdsourcing. In Wired
magazine, 14: 1-4.
Jia, Y., Wang, Y., Cheng, X., Jin, X., and Guo, J.
(2014). Openkn: An open knowledge compu-
tational engine for network big data. In Pro-
ceedings of Advances in Social Networks Anal-
ysis and Mining (ASONAM).
Jia, Y., Wang, Y., Lin, H., Jin, X., and Cheng, X.
(2016). Locally adaptive translation for knowl-
edge graph embedding. In Proceedings of the
Thirtieth AAAI Conference on Artificial Intelli-
Jin, Z., Cao, J., Zhang, Y., and Luo, J. (2016).
News verification by exploiting conflicting so-
cial viewpoints in microblogs. In Proceedings of
the Thirtieth AAAI Conference on Artificial In-
telligence (AAAI’16).
Liu, D., Wang, Y., Jia, Y., Li, J., and Yu, Z. (2014).
Lsdh: a hashing approach for large-scale link
prediction in microblogs. In Twenty-Eighth
AAAI Conference on Artificial Intelligence.
Long, Y., Lu, Q., Xiang, R., Li, M., and Huang,
C. (2017). Fake news detection through multi-
perspective speaker profiles. In Proceedings of
The International Joint Conference on Natural
Language Processing.
McIntire, G. (2018). Fake real news dataset. In
GeorgeMcIntire’s Github.
Mieghem, V. (2011). Human psychology of common
appraisal: The reddit score. In IEEE Transac-
tions on Multimedia, 13: 1404-1406.
Miller, G. A. (1995). Wordnet: a lexical database
for english. In Communications of the ACM
Pan, J. Z., Pavlova, S., Li, C., Li, N., Li, Y., and Liu,
J. (2018). Content based fake news detection us-
ing knowledge graphs. In International Seman-
tic Web Conference (ISWC), Monterey, Califor-
nia, volume 11136 of Lecture Notes in Computer
Science, pages 669–683, Cham. Springer Inter-
national Publishing.
Rashkin, H., Choi, E., Volkova, S., Jang, J. Y., and
Choi, Y. (2017). Truth of varying shades: Ana-
lyzing language in fake news and political fact-
checking. In Proceedings of the 2017 Confer-
ence on Empirical Methods in Natural Language
Rubin, V. L., Chen, Y., and Conroy, N. J. (2015).
Deception detection for news: three types of
fakes. In Proceedings of the 78th ASIS&T An-
nual Meeting.
Ruchansky, N., Seo, S., and Liu, Y. (2017). Csi: A
hybrid deep model for fake news detection. In
Proceedings of the 2017 ACM Conference on In-
formation and Knowledge Management.
Shao, C., Ciampaglia, G. L., Flammini, A., and
Menczer, F. (2016). Hoaxy: A platform for
tracking online misinformation. In Proceedings
of the 25th International Conference Companion
on World Wide Web.
Tacchini, E., Ballarin, G., Vedova, M. L. D., Moret,
S., and Alfaro, L. (2017). Some like it hoax: Au-
tomated fake news detection in social networks.
In Proceedings of the Second Workshop on Data
Science for Social Good.
Volkova, S., Shaffer, K., Jang, J. Y., and Hodas, N.
(2017). Separating facts from fiction: Linguistic
models to classify suspicious and trusted news
posts on twitter. In Proceedings of the 55th An-
nual Meeting of the Association for Computa-
tional Linguistics.
Wang, W. Y. (2017). Liar, liar pants on fire: A new
benchmark dataset for fake news detection. In
arXiv:1705.00648 [cs.CL].
Wu, W., Li, H., Wang, H., and Zhu, K. Q. (2012).
Probase: A probabilistic taxonomy for text un-
derstanding. In Proceedings of the 2012 ACM
SIGMOD International Conference on Manage-
ment of Data, 481-492.
ResearchGate has not been able to resolve any citations for this publication.
Conference Paper
Full-text available
In recent years, the reliability of information on the Internet has emerged as a crucial issue of modern society. Social network sites (SNSs) have revolutionized the way in which information is spread by allowing users to freely share content. As a consequence, SNSs are also increasingly used as vectors for the diffusion of misinformation and hoaxes. The amount of disseminated information and the rapidity of its diffusion make it practically impossible to assess reliability in a timely manner, highlighting the need for automatic hoax detection systems. As a contribution towards this objective, we show that Facebook posts can be classified with high accuracy as hoaxes or non-hoaxes on the basis of the users who "liked" them. We present two classification techniques, one based on logistic regression, the other on a novel adaptation of boolean crowdsourcing algorithms. On a dataset consisting of 15,500 Facebook posts and 909,236 users, we obtain classification accuracies exceeding 99% even when the training set contains less than 1% of the posts. We further show that our techniques are robust: they work even when we restrict our attention to the users who like both hoax and non-hoax posts. These results suggest that mapping the diffusion pattern of information can be a useful component of automatic hoax detection systems.
Full-text available
We propose the first multistage intervention framework that tackles fake news in social networks by combining reinforcement learning with a point process network activity model. The spread of fake news and mitigation events within the network is modeled by a multivariate Hawkes process with additional exogenous control terms. By choosing a feature representation of states, defining mitigation actions and constructing reward functions to measure the effectiveness of mitigation activities, we map the problem of fake news mitigation into the reinforcement learning framework. We develop a policy iteration method unique to the multivariate networked point process, with the goal of optimizing the actions for maximal total reward under budget constraints. Our method shows promising performance in real-time intervention experiments on a Twitter network to mitigate a surrogate fake news campaign, and outperforms alternatives on synthetic datasets.
Conference Paper
Full-text available
Tabloid journalism is often criticized for its propensity for exaggeration, sensationalization, scare-mongering, and otherwise producing misleading and low quality news. As the news has moved online, a new form of tabloidization has emerged: ‘clickbaiting.’ ‘Clickbait’ refers to “content whose main purpose is to attract attention and encourage visitors to click on a link to a particular web page” [‘clickbait,’ n.d.] and has been implicated in the rapid spread of rumor and misinformation online. This paper examines potential methods for the automatic detection of clickbait as a form of deception. Methods for recognizing both textual and non-textual clickbaiting cues are surveyed, leading to the suggestion that a hybrid approach may yield best results.
Conference Paper
The topic of fake news has drawn attention both from the public and the academic communities. Such misinformation has the potential of affecting public opinion, providing an opportunity for malicious parties to manipulate the outcomes of public events such as elections. Because such high stakes are at play, automatically detecting fake news is an important, yet challenging problem that is not yet well understood. Nevertheless, there are three generally agreed upon characteristics of fake news: the text of an article, the user response it receives, and the source users promoting it. Existing work has largely focused on tailoring solutions to one particular characteristic which has limited their success and generality. In this work, we propose a model that combines all three characteristics for a more accurate and automated prediction. Specifically, we incorporate the behavior of both parties, users and articles, and the group behavior of users who propagate fake news. Motivated by the three characteristics, we propose a model called CSI which is composed of three modules: Capture, Score, and Integrate. The first module is based on the response and text; it uses a Recurrent Neural Network to capture the temporal pattern of user activity on a given article. The second module learns the source characteristic based on the behavior of users, and the two are integrated with the third module to classify an article as fake or not. Experimental analysis on real-world data demonstrates that CSI achieves higher accuracy than existing models, and extracts meaningful latent representations of both users and articles.
Automatic fake news detection is a challenging problem in deception detection, and it has tremendous real-world political and social impacts. However, statistical approaches to combating fake news has been dramatically limited by the lack of labeled benchmark datasets. In this paper, we present liar: a new, publicly available dataset for fake news detection. We collected a decade-long, 12.8K manually labeled short statements in various contexts from, which provides detailed analysis report and links to source documents for each case. This dataset can be used for fact-checking research as well. Notably, this new dataset is an order of magnitude larger than previously largest public fake news datasets of similar type. Empirically, we investigate automatic fake news detection based on surface-level linguistic patterns. We have designed a novel, hybrid convolutional neural network to integrate meta-data with text. We show that this hybrid approach can improve a text-only deep learning model.
Cognitive computing systems require human labeled data for evaluation, and often for training. The standard practice used in gathering this data minimizes disagreement between annotators, and we have found this results in data that fails to account for the ambiguity inherent in language. We have proposed the CrowdTruth method for collecting ground truth through crowdsourcing, that reconsiders the role of people in machine learning based on the observation that disagreement between annotators provides a useful signal for phenomena such as ambiguity in the text. We report on using this method to build an annotated data set for medical relation extraction for the $cause$ and $treat$ relations, and how this data performed in a supervised training experiment. We demonstrate that by modeling ambiguity, labeled data gathered from crowd workers can (1) reach the level of quality of domain experts for this task while reducing the cost, and (2) provide better training data at scale than distant supervision. We further propose and validate new weighted measures for precision, recall, and F-measure, that account for ambiguity in both human and machine performance on this task.