Available via license: CC BY 4.0
Content may be subject to copyright.
arXiv:2505.10746v1 [cs.CY] 15 May 2025
ChestyBot: Detecting and Disrupting Chinese
Communist Party Influence Stratagems
Matthew Stoffolano,Ayush Rout,Justin M. Pelletier
Rochester Institute of Technology
Email: jxpics@rit.edu
“We’re surrounded. That simplifies the problem.” -Chesty Puller
Abstract—Foreign information operations conducted by Rus-
sian and Chinese actors exploit the United States’ permissive
information environment. These campaigns threaten democratic
institutions and the broader Westphalian model. Yet, existing
detection and mitigation strategies often fail to identify active
information campaigns in real time. This paper introduces
ChestyBot, a pragmatics-based language model that detects
unlabeled foreign malign influence tweets with up to 98.34%
accuracy. The model supports a novel framework to disrupt
foreign influence operations in their formative stages.
Index Terms—social cybersecurity, information warfare, influ-
ence stratagems, NLP pragmatics
I. INTRODUCTION
Foreign influence campaigns—particularly those
attributed to Russia during the 2016 U.S. Presidential
Election—demonstrated how state-sponsored social media
operations can destabilize democratic societies [1]. During
that campaign, social media posts emanating from one state
– Russia – probably represented an intentional effort to
influence the internal affairs of another country – the United
States. Though these efforts may not have changed election
outcomes, they nonetheless constitute an erosion of the
Westphalian state model itself [2].
In recent years, China has attempted to use social media
to influence foreign perceptions of internal matters such as
the Beijing 2022 Winter Olympics, the origins of COVID-19,
and the human rights abuses in Xinjiang [3]. Despite these
initiatives, China has (as far as we can tell at the time of this
writing) not performed a successful large-scale disinformation
campaign directed against U.S. internal interests. However, in
recent years China’s attacks have adopted similar mannerisms
to Russia [4], which could increase the risk of impact.
Standard ways to detect influence campaigns include look-
ing for bot networks [5], observing coordinating behavior [6],
and implementing various types of retrospective analysis [7].
Though useful in retrospect, most methods struggle to detect
campaigns as they unfold. Early detection of foreign malign
influence is important because taking preventive measures in
the early stages of an influence campaign on social media
can permanently stunt the campaign’s growth [8]. Improving
the current methods requires looking at a combination of the
attacker methods and victim communities.
To better understand the trajectory of this threat and the
potential for foreign malign influence to impact U.S. internal
affairs, this project aims to provide a proof of concept for
detecting the influence campaigns undertaken by the Chinese
Communist Party(CCP)/People’s Liberation Army(PLA) in its
earliest stages. We, therefore, propose a novel method to detect
four influence stratagems used by the CCP, with data generated
during the 2022 U.S. Mid-term Elections.
II. BACKGROU ND
A. Social Cybersecurity
Social science seeks to understand how innate human traits
affect the world [9]. In recent years, researchers have leveraged
computational social science, societal computing, data science,
and policy studies in the emergent field called Social Cyberse-
curity [10]. This field considers linguistic, business, group, and
behavioral analyses using techniques such as dating mining or
forensics [7].
B. Information Warfare
Information Warfare manages the narratives that reach a
population to achieve decisive ideological outcomes. This can
be done by dropping pamphlets over a specific region or
spreading disinformation on social media [11]. The strategies
used against the United States are visualized below in Figure
1.
C. Influence Stratagems
Influence Campaigns consist of multiple operations employ-
ing various tactical-level strategies (called stratagems through-
out this paper for clarity). These stratagems are probably best
described by The Taxonomy of Influence Strategies, provided
as an artifact of a NATO-funded examination of Russia’s
annexation of the Crimean Peninsula [12], depicted in Figure
4.
Though Russia is more stentorian than China in foreign ma-
lign influence directed against the United States, we stipulate
that there are likely overlaps between the Russian propaganda
playbook and what China is likely to use. We discuss this
more in section IV.
D. Echo Chambers
Implementing algorithms on social media to improve the
user experience has created a phenomenon known as echo
chambers. Members of these isolated communities interact
mostly or exclusively with other community members, with
few or no outside influences. Mapping these communities
Fig. 1: Each colored box represents a decision that can be influenced by the application of Stratagems. The introduction of
stimuli prompts the actor to make a decision beneficial to an influencer’s selected course of action [Starting with decision I2].
The deliberate application of Stratagems directs decisions toward a specific outcome [Green Arrow to A5] and, failing that,
accounts for the range of possible decisions that would lead to acceptable managed outcomes [Green Boxes]. Adapted from
communications with Colonel Drew Cukor, USMC in 2019.
shows how information flows in the online fora they frequent.
Figure 6 depicts the flow of information between two chambers
through an intermediary (a liminal node), which represents
high betweenness centrality. This measure of influence is an
important component of identifying relationships among echo
chambers in social networks [13].
III. REL ATED WORK
Several attempts to analyze and detect disinformation exist.
It is beyond the scope of this paper to provide an exhaustive
listing. One important exemplar of this body of knowledge,
conducted by researchers at Lincoln Labs, focuses on finding
disinformation and identifying critical actors. Their method-
ology focuses on starting with a keyword search to build a
narrative [14]. This in part informed our approach to look
for keywords that could indicate an influence stratagem in the
datasets that aligned with previous methods used by the CCP
to validate ChestyBot’s training model.
Each nation, however, does have variation in how they
conduct influence campaigns. During the Covid pandemic,
the CCP employed operations to spread their narrative by
trying to redirect through sharing positive stories and recast
through rewriting history.[15] Further, in further events such
as the Beijing protest, the CCP showed more strategies by
seeking to inform the global population with information
meant to affect the public’s view of the events. They would
also attempt to deflect events by framing the protesters in
morally compromising positions, such as spreading stories of
them using M320 a tool used in the United States military for
employing High Explosives. All of these narratives are pushed
with the idea of spreading the Chinese stories.[16]
Also, the analysis of echo chambers has been performed by
many other researchers seeking to understand how misinfor-
mation spreads. The most common focus reveals indicators
of a bad actor releasing misinformation with the intent of
others picking it up and spreading it. Sometimes this is done
in conjunction with social media bots that amplify the mis-
information [17]. Though our approach is similar, we extend
this pursuit by looking for detection of influence campaigns,
which can contain misinformation, disinformation, and even
true information that is framed in a specific way.
Finally, a 2015 U.S. Special Operations Command publi-
cation titled ”Operating in the Human Domain” specifically
highlighted the difficulty of measuring information [18]. The
same publication also featured the ease of manipulating in-
formation as a fundamental challenge of achieving desired
affects. This observation in particular informs our approach
to frame the problem we seek to solve as one that improves
the measurability of information operations. This approach is
depicted in Figure 2.
Fig. 2: Our main effort is to make the external measura-
bility of informational artifacts easier, which is a problem
well described in a 2015 U.S. Special Operations Command
publication [18].
IV. MET HO D
We employed the snowball method, starting with accounts
with known ties to the CCP and adding identified aligned
accounts (layer one) as well as another layer of accounts
(layer two) who are aligned with layer one accounts. We
consider this snowball to be a representative sample of an
echo chamber. We also employed the Louvain Algorithm to
identify echo chambers among a wider set of data for testing
purposes. After this sampling, we employed a natural language
process to detect possible propaganda tweets according to the
four specific influence strategies (stratagems): inform, invoke,
deflect, and recast. The remainder of this section describes
each component of this method in detail.
A. Platform Selection
Data was collected on X (Twitter) over other platforms
because of the data type, access, and tools. Any platform fo-
cused on images or videos was excluded to remove complexity
and focus on text communication. Text-based communication
was further narrowed down based on site allowance to give
researchers access to data and a tool that could implement with
the collection[19]. Combined with Tweepy[20] and research
access to X (Twitter), we were able to pull down current user
information to build data sets for training and testing.
B. Snowball Sampling Method (SSM)
Discovering any group that is trying to employ an influ-
ence campaign is met with an inherent problem due to the
clandestine nature of the operation. Thus, it became necessary
to employ the snowball sampling method(SSM) to collect
the necessary data to research operatives employing these
campaigns. SSM emerged as a useful method for finding hard-
to-reach groups[21], [22] and is used by anthropologists to
locate groups in a conflict environment where inherent mistrust
exists[23]. Further, SSM has also been implemented on social
media to find political dynamics in unitary countries[24]. SSM
is ideal for characterization of influence behaviors emanating
from known or suspected influencers because it has been
widely used in behavioral research[25], [26] and remains
one of the most popular methods of sampling in qualita-
tive research that examine characteristics of networking and
referral[22], [27].
SSM makes it possible to create a network containing
accounts suspected of involvement in an influence campaign.
Each echo chamber shows user account interactions and the
proliferation of ideas across the echo chamber. In this case,
SSM began with identifying known CCP accounts and per-
forming a recursive search on friends, followers, and retweets.
In aggregate, this allowed us to build a list of potential
accounts that could be involved in an influence campaign
emanating from one actor ”Jane Doe”. Using the results of
SSM, we constructed a data set of potential influence activities,
(Set1), described below.
C. Community Identification and Edge Weighting
The Louvain Algorithm was implemented on Set2 to
map out the echo chambers. The Algorithm originated in
2008 and has emerged as the de-facto standard for quickly
extracting community structure from large networks [28]. The
Louvain Algorithm is a heuristic method based on modularity
optimization that analyzes large web graphs to identify com-
munities. It works in two phases: (1) modularity optimization
through local community changes and (2) aggregation of
communities to build a new network of communities. These
phases repeat until no modularity gain is possible. We utilized
the Louvain Algorithm because of its ability to map out vast
online communities in a short period of time, as well as its
ability to identify the artificial nodes [29], and the influential
nodes in a network [30]. This allowed for quickly assembling
Fig. 3: We used the Snowball Sampling Method to build Set1.
the echo chambers and being able to understand the role of
nodes in the community.
We optimized the communities by defining edge weights,
where likes received a weight of one and retweets and
friends/followers received a weight of 10.
We optimized the communities by defining edge weights,
where likes received a weight of one and retweets and
friends/followers received a weight of 10.
D. Data Retrieval and Anonymization
Two datasets were collected (Set1 and Set2) using differ-
ent methods based on the SSM. Both Sets contained a hash
of the user ID, creation date, and tweet contents.
Set1 started with Jane Doe, a CCP member who com-
monly pushes influences narratives. Pulling her tweets from 1
October 2022 to 8 November 2022 provided the information to
construct the first layer using SSM. For each of Jane’s tweets,
we collected 20 unique random users who retweeted the tweet.
The new accounts comprise the first layer, and then the same
process of tweet search is conducted on the recently collected
accounts to form the second and final layers. In total, there
were approximately 10,000 tweets collected, a majority of
those were posted in the English language. This process is
depicted in summary in Figure 3.
Set2 started with a list of CCP members, including
Jane Doe, and extended to retrieve data on friends/followers,
retweets, and likes. Using the Louvain Algorithm, we assigned
each account to an echo chamber, as depicted in Figure 6. Then
one of the echo chambers was randomly selected to collect
tweets from 1 January 2010 to 25 February 2023. This resulted
in a dataset of approximately 280,000 tweets.
E. Data Labeling
Both Set1 and Set2 data sets initially contained a hashed
user ID, the creation date, and a tweet. Set1 was used for
training and therefore encoded according to four influence
stratagems: inform, invoke, deflect, and recast. These influence
strategems were identified and described in a NATO-funded
taxonomy[12] Though Russia is probably more likely to use a
different range of influence stratagems than China, we believe
that Chinese Communist Party members are also likely to
use the four highlighted stratagems (inform,invoke,deflect,
recast) from the Russian propaganda playbook.
Set2 was used for testing. Encoding was done manually
where each tweet was labeled according to the presence of
any of the four stratagems. If we found the tweet matched
one of the stratagems, it would be marked as true. A tweet
only needed to have one true strategem to become classified
as a tweet that was part of an influence campaign.
Each of these sets, and the influence stratagems themselves,
are defined in the remainder of this sub-section. The entire
taxonomy of influence stratagems and our highlighted
selections are depicted below in Figure 4.
Set1
Inform and invoke focus on framing the debate.
Inform focuses on educating users about specific infor-
mation. One prominent example we found is the Nordstrom
Pipeline Leak, as in ”#ChinaDailyCartoon Who gains most
from #NordStream sabotage?”
Invoke focuses on connecting to a particular event. One of
the invoked tweets used in training is ”US shames MS, while
ignores its own vile human rights records -George Floyd...”.
Deflect and recast are diversion tactics, where attackers attempt
to spin a conversation.
A common form of deflection is bringing up the US
treatment of minorities during international human rights
discussions. An example deflect tactic we observed in Set1
stated: ”Firearm, abortion, economy, immigration ... all be-
come ammunition for U.S. partisan fight, a fight making the
house of American politics totter”.
Recasting is simply another term for rebranding. For
example, an example recast tactic we observed in the
snowball stated: ”NATO is not a defense alliance, it’s a war
machine. Ask the people of Afghanistan, Iraq or Libya”.
If the tweet was determined to have one of these four
stratagems, it was marked as a propaganda tweet. If not, we
encoded it as not a propaganda tweet. This led to 882 total
tweets encoded, with 62 of those designated as propaganda.
Set2
Set2 was encoded using dates and keywords; each set of
dates was a time that a campaign had been expected to have
occurred. Though useful in concept, we found the list of
keywords to be vastly inferior as an indicator of campaign
following the building and testing of ChestyBot.
F. Building ChestyBot
Through classifying small data using known influence
stratagems, a detector-bot using machine learning can detect
potential tweets of a campaign. Once one tweet is detected,
Fig. 4: The NATO-funded taxonomy [12] depicted here provides a digest of Russian influence stratagems used during the
annexation of Crimea.
a defender can further inspect the area where the campaign
occurs. Improving early detection and tracking campaign
contagion provides essential tools for studying orchestrated
influence campaigns. This is a critical task for our effort.
There exist promising models that might help automate the
identification of influence stratagems within tweets, as in Ten-
sor Flow’s natural language processing(NLP) toolset[31]. In
particular, we considered two pre-trained models (ROBERTO
and BERT) but found them incompatible. These tools and
many others like them have an over-emphasis on sentiment
analysis in text prediction but do not explicitly consider the
logical rationale or context employed within the language
[32], [33]. The logical rationale and context – what someone
means – is called pragmatics, while the emotion behind
the words – how someone feels – is called semantics. Our
observation of this limitation among available models largely
conforms to previous findings of Li, Thomas, and Liu, who
call on academic researchers and NLP developers to place
less emphasis on semantics and more on pragmatics [34].
Therefore, instead of using pre-trained models, we trained a
pragmatics-focused NLP model (ChestyBot) using Set1 and
tested it using Set2.
We chose a convolutional neural network (CNN) for this
task of text classification, rather than a traditional feedforward
neural network, because CNNs are better able to capture
the local patterns and dependencies in the input text data
[35]. Text data exhibits a strong sequential dependency, where
the meaning of a word is shaped by the context in which
it appears. Traditional feedforward neural network models,
which treat input data as a fixed-size vector, fail to fully
account for these contextual dependencies. Specifically, the
feedforward network cannot distinguish between the different
senses of a given the word when used in various contexts. On
the other hand, a CNN is able to capture the local patterns
in the input data through the use of convolutional filters. In
the context of text classification, the filters can be designed
to recognize common n-grams (i.e., contiguous sequences
of n words) in the input text data. The pooling layers can
then be used to downsample the output of the convolutional
layers, which reduces the dimensionality of the input data
while preserving the most salient features [36]. In addition
to capturing local patterns, a CNN can also learn hierarchical
representations of the input data. Each layer in the CNN learns
progressively more complex representations of the input data,
allowing the model to capture both low-level features such
as individual words, as well as higher-level features such as
phrases or sentences.
For the purposes of this investigation, we built a CNN archi-
tecture using TensorFlow’s Keras library. The input data con-
sisted of a set of text strings from Set1, which were encoded
as one-hot vectors of length ’input dim’, where input dim was
set to 1536 which represents the size of the vocabulary used to
encode the text. One hot encoding is a way of numerically rep-
resenting text where each word in the vocabulary is assigned
a unique integer index [37]. Each element of the vector is set
to zero, except for the element corresponding to the index of
the word in the vocabulary, which is set to one. The resulting
data is then padded to a fixed length of input length using the
tf.keras.preprocessing.sequence.pad sequences() function. The
encoded and padded data is then passed through an embedding
layer, which maps each index to a dense vector representation
of length dense vectors.
The output of the embedding layer is a 3D tensor of
shape (num examples, input length, dense vectors), where
each element corresponds to the embedding of a word. Fol-
lowing the embedding layer, a 1D convolutional layer with
32 filters of size 5 is applied to the embeddings. This layer
performs a convolution operation along the temporal dimen-
sion of the input, capturing local patterns in the sequence
of embeddings. The output of this layer is a 3D tensor of
shape (num examples, new input length, num filters), where
new input length = input length - 4 due to the size of
the filters used. Subsequently, a 1D max pooling layer is
applied to the output of the convolutional layer, which reduces
the dimensionality of the output by selecting the maximum
value over a window of size 2. The resulting output is
a 3D tensor of shape (num examples, new input length/2,
num filters). At this point, the output of the max pooling
layer is flattened using the tf.keras.layers.Flatten() layer. This
operation reshapes the output into a 2D tensor of shape
(num examples, new input length/2 * num filters), where
each row corresponds to an example in the batch and each
column corresponds to a flattened feature. Finally, the flattened
output is passed through a dense layer with a single output unit
and a sigmoid activation function, which produces a binary
classification output. In simple terms, this classification output
was ”propaganda” or ”not propaganda”, based on the influence
stratagem encodings in Set1.
Ultimately, this led to the creation of ChestyBot, which
revealed efficient training (depicted in Figure 5) and promising
test performance (described in the next section).
Fig. 5: Training for ChestyBot
V. FINDINGS
Our research focused on finding methods to detect a cam-
paign in the early stages when the impact is still containable.
Using SSM, we found it is possible to discover previously
undetected accounts that were part of a campaign. Also, by
using data generated by these accounts, we were able to map
communities and determine how each one was connected.
Together, this revealed the liminal nodes that may share
information across other echo chambers. We were also able
to apply a novel pragmatics-based NLP model (ChestyBot)
that was able detect a campaign in one of these chambers.
The remainder of this section describes more fully our proof
of concept model.
ChestyBot performed well in identifying influencing
tweets; during training, it had a validation accuracy of 99.44
percent. Testing it against a random echo chamber in Set2,
ChestyBot found 241 potential propaganda tweets. 92.53
percent of those were True Positive (Obvious True), 5.81
percent of tweets were accurately classified according to the
context (Context-Dependent True), and 1.66 percent false
positives. Of note, though it would benefit from fine-tuning
to increase the likely hood of detecting tweets, fine-tuning to
catch all propaganda tweets is probably unnecessary because
our objective is reliable detection of ongoing campaigns
rather than every part of every campaign. Even so, ChestyBot
could be improved through additional data labeling and
training, which would allow for improved indicators and
warning of a campaign underway. These improvements are
likely to minimize false positives that occur from an abstract
connection.
Obvious True:
•”Tariffs are not the right way to go. American businesses
and consumers want solutions, not sanctions”
•”It’s not just your local supermarket. Eggs are 60 percent
more expensive than last year in the U.S”
•”The issues related to Xinjiang & Hong Kong are not
about democracy, human rights, ethnicity or religion,
but about anti-terrorism, anti-secession, about protecting
people’s lives, safeguarding China’s national sovereignty,
security, and development interests”.
•”China and the US do compete. But how should the
competition play out? Our relations should not be like
the intensely confrontational American football match.
There should be no offensive team or defensive team,
no touchdown, no quarterback sack.”
•”59 people have been killed in 36 mass shootings in the
U.S. since the start of the year.”
Context-Dependent True:
•”How a country is doing on human rights is essentially
gauged by whether the interests of its people are upheld,
and whether they enjoy a growing sense of fulfillment,
happiness and security”
False Positive:
•”The United Nations report on the PRC’s human rights
violations against the Uyghur people and other minority
groups in Xinjiang is clear and devastating. The U.S. calls
on Beijing to cease these atrocities”
•“Everyone thought the billionaires could save us —
clearly, they’re not engaged,” grumbled one executive at
a billionaire-owned newsroom. “It’s not a good model
because they lose interest or they get pissy.”
VI. A NOVE L FRA ME WO RK F OR IN FLU ENCE OPERATIO N
DISRUPTION
The automated detection of malign influence campaigns
enables a more proactive approach to social cybersecurity,
particularly when combined with network analysis techniques.
Once an influence campaign is detected using tools such as
ChestyBot, a key step in disrupting its efficacy is to identify
the liminal nodes—users or accounts that bridge otherwise
distinct echo chambers.
Fig. 6: We incorporated the Louvain Algorithm to build
echo chambers and identify liminal nodes between multiple
chambers.
These liminal nodes exhibit high betweenness centrality,
meaning they occupy structurally important positions in the
social graph that allow them to serve as information conduits
between insular communities. Messages originating from a
malign actor and reaching these nodes are more likely to
“break out” of their original echo chamber and achieve broader
dissemination.
By coupling early detection of propaganda tweets with
echo chamber community mapping (e.g., via the Louvain
Algorithm), we can observe where influence operations are
gaining traction and predict their breakout potential. When a
malign message is picked up by a liminal node, that message
becomes a candidate for targeted disruption—such as counter-
messaging, deprioritization in feeds, or account moderation.
Operationally, this means that automated detection systems
should not only flag malicious content but also evaluate
the network position of users who engage with it. A tweet
that reaches a highly central liminal node may pose more
strategic risk than one that circulates exclusively within a
single community.
This framework transforms detection from a passive, foren-
sic activity into an active defense mechanism capable of
shaping the informational battlespace in real time. As depicted
in Figure 6, the integration of message classification and
network topology offers a dynamic framework for mitigating
malign influence at the point of maximum potential impact.
VII. CONCLUSIONS AND DISCUSSION
The connection of social media has created a relatively
new domain for information warfare. Around the world, many
countries are deploying varying strategies to win the hearts
and minds of foreign populations. However, many strategies
in this domain remain underdeveloped possibly due to an over-
focus on technology and retrospective analysis. Our approach
focuses on how information flows in the human domain, and
we suggest that engaging with these technologies can facil-
itate a better understanding of and defense against influence
campaigns.
The Snowball Sampling Method allows for finding accounts
that may participate in a campaign and mapping the communi-
ties around these accounts. This allowed for the deployment of
ChestyBot, which can detect potential campaign tweets based
on the stratagems that an attacker employs. The ability to
detect these tweets during a live campaign allows information
that would otherwise remain hidden from other researchers.
This can allow for early campaign detection.
Furthermore, the Louvain Algorithm allows the isolation of
specific communities and the identification of liminal nodes
connecting echo chambers. Each of these liminal nodes is a
weak point in an adversary’s campaign; it could represent a
sort of key terrain. We propose that facilitating effects against
those nodes probably provides a way to manage information
flow between online communities.
By combining these methods, we were able to provide
a proof of concept mechanism – ChestyBot – capable of
detecting influence campaigns emanating from the Chinese
Communist Party. Furthermore, we propose that such a de-
tection mechanism can inform follow-on effects (ie. block,
canalize, contain, defeat, destroy, disrupt, fix, interdict, iso-
late, neutralize, suppress, turn) on foreign malign influence
campaigns. We believe our proof of concept and the effects
it allows could help re-gain the initiative in the information
sphere.
VIII. LIMITATIONS AND FUTURE WORK
Though X (Twitter) provides many advantages, retrospective
analyses are limited due to limitations on the information that
is accessible/collected. This impacted our ability to build a
complete scenario. For example, we could not see when one
account became another’s friend/follower, so the communities
collected may have accounts that were community members
during the campaign. Furthermore, our access to seeing likes
was limited to only the most recent 100. Even though we
believe the effects from this temporal bias were minimized by
prioritizing edge weights for retweets and following activities,
we believe a more robust real-time data collection could
improve echo chamber construction.
Two recommended advances for ChestyBot are to (1) im-
prove accuracy and (2) to generalize detection capabilities. To
improve accuracy, we point out that Set1 data only contained
62 tweets labeled as propaganda. Though ChestyBot per-
formed well (better than expected, really), a more substantial
data set would almost certainly improve the ability to detect
different forms of the trained influence stratagems. Further,
ChestyBot was only trained on 4 stratagems that would match
likely Chinese methods. Including more data and exploring
ChestyBot’s capability to detect other stratagems would make
it more useful in the field. Thogether, these improvements
would probably make our proof of concept model better able
to detect foreign malign influence campaigns emanating from
other actors across the competition continuum.
Finally, there remain important and necessary ethical con-
siderations regarding the weaponization of ChestyBot and the
associated influence operation disruption framework. Many
major systems will have something to say about this. We
consider only a few here as fodder for future inquiry. From
a Catholic ethical standpoint, the inherent dignity of every
human person demands caution when deploying technologies
that may surveil or manipulate information environments,
even in the name of defense[38]. Classical liberal democratic
norms similarly require that interventions in public discourse
respect individual autonomy, freedom of expression, and the
pluralism essential to a free society[39], [40]. Yet from a
military operational perspective, the growing sophistication of
adversarial influence operations necessitates proactive capa-
bilities that can detect and neutralize emerging threats in the
information domain. These tensions reveal a critical tradeoff:
the very tools that safeguard national sovereignty and civilian
morale may, if ungoverned or misapplied, erode the moral
and civic foundations they aim to protect. Therefore, any
operational use of such systems must be accompanied by
rigorous oversight, strict proportionality, clear attribution, and
a reaffirmed commitment to both just war principles and the
preservation of open democratic dialogue.
IX. ACK NOWLEDGEMENTS
We would like to acknowledge the support and contributions
from CPT Noah Demoes and entire the team at CyberRecon.
J.M. Pelletier would like to express deep gratitude for the
ongoing support he receives from the Ordo Praedicatorum.
REFERENCES
[1] A. Badawy, A. Addawood, K. Lerman, and E. Ferrara, “Characterizing
the 2016 russian ira influence campaign,” Social Network Analysis and
Mining, vol. 9, pp. 1–11, 2019.
[2] J. Lukito, “Coordinating a multi-platform disinformation campaign:
Internet research agency activity on three us social media platforms,
2015 to 2017,” Political Communication, vol. 37, no. 2, pp. 238–255,
2020.
[3] C. S. Jacobs and K. M. Carley, “# whodefinesdemocracy: Analysis on a
2021 chinese messaging campaign,” in Social, Cultural, and Behavioral
Modeling: 15th International Conference, SBP-BRiMS 2022, Pittsburgh,
PA, USA, September 20–23, 2022, Proceedings. Springer, 2022, pp.
90–100.
[4] M. Repnikova and S. Bret, “How the people’s republic of china amplifies
russian disinformation,” US Department of State. Accessed April, vol. 25,
p. 2022, 2022.
[5] D. Pacheco, P.-M. Hui, C. Torres-Lugo, B. T. Truong, A. Flammini, and
F. Menczer, “Uncovering coordinated networks on social media,” arXiv
preprint arXiv:2001.05658, vol. 16, 2020.
[6] L. Vargas, P. Emami, and P. Traynor, “On the detection of disinformation
campaign activity with network analysis,” in Proceedings of the 2020
ACM SIGSAC Conference on Cloud Computing Security Workshop,
2020, pp. 133–146.
[7] K. M. Carley, “Social cybersecurity: an emerging science,” Computa-
tional and mathematical organization theory, vol. 26, no. 4, pp. 365–
381, 2020.
[8] S. Lewandowsky and S. Van Der Linden, “Countering misinformation
and fake news through inoculation and prebunking,” European Review
of Social Psychology, vol. 32, no. 2, pp. 348–384, 2021.
[9] R. A. Hinde, Individuals, relationships and culture: Links between
ethology and the social sciences. CUP Archive, 1987.
[10] National Academies of Sciences, Engineering, and Medicine, A Decadal
Survey of the Social and Behavioral Sciences: A Research Agenda for
Advancing Intelligence Analysis. Washington, DC: National Academies
Press, 2019.
[11] J. P. Farwell, Information warfare forging communication strategies for
twenty-first century operational environments. Marine Corps University
Press, Quantico, VA, USA, 2020.
[12] A. Kelly and C. Paul, DECODING CRIMEA. PINPOINTING THE
INFLUENCE STRATEGIES OF MODERN INFORMATION WARFARE.
NATO Strategic Communications Centre of Excellence, 2020.
[13] N. Kratzke, “How to find orchestrated trolls? a case study on identifying
polarized twitter echo chambers,” Computers, vol. 12, no. 3, p. 57, 2023.
[14] S. T. Smith, E. K. Kao, E. D. Mackin, D. C. Shah, O. Simek, and
D. B. Rubin, “Automatic detection of influential actors in disinformation
networks,” Proceedings of the National Academy of Sciences, vol. 118,
no. 4, p. e2011216118, 2021.
[15] V. Molter and R. DiResta, “Pandemics & propaganda: How chinese
state media creates and propagates ccp coronavirus narratives,” Harvard
Kennedy School Misinformation Review, vol. 1, no. 3, 2020.
[16] R. DiResta, C. Miller, V. Molter, J. Pomfret, and G. Tiffert, Telling
China’s story: the Chinese communist party’s campaign to shape global
narratives. Stanford Internet Observatory Stanford, CA, 2020.
[17] C. Diaz Ruiz and T. Nilsson, “Disinformation and echo chambers:
How disinformation circulates on social media through identity-driven
controversies,” Journal of Public Policy & Marketing, vol. 42, no. 1, pp.
18–35, 2023.
[18] U. S. O. Command, “Operating in the human domain,” Tampa: USSO-
COM, 2015.
[19] S. J. Chaudhary and S. Niveditha, “Twitter sentiment analysis using
tweepy,” Twitter Sentiment Analysis using Tweepy, vol. 8, no. 04, p.
4512, 2021.
[20] J. Roesslein, “tweepy documentation,” Online] http://tweepy. readthe-
docs. io/en/v3, vol. 5, p. 724, 2009.
[21] L. A. Goodman, “Snowball sampling,” The annals of mathematical
statistics, pp. 148–170, 1961.
[22] C. Parker, S. Scott, and A. Geddes, “Snowball sampling,” SAGE research
methods foundations, 2019.
[23] N. Cohen and T. Arieli, “Field research in conflict environments:
Methodological challenges and snowball sampling,” Journal of peace
research, vol. 48, no. 4, pp. 423–435, 2011.
[24] J. A. Clark, “Field research methods in the middle east,” PS: Political
Science & Politics, vol. 39, no. 3, pp. 417–424, 2006.
[25] T. Yoshida, T. Gotoh, N. Tomizawa, and K. Ikeda, “Snowball sampling
consumer behaviour research to characterise the influence of market
mavens on social networks,” International Journal of Intelligent Systems
Technologies and Applications, vol. 12, no. 3-4, pp. 268–282, 2013.
[26] C. Noy, “Sampling knowledge: The hermeneutics of snowball sampling
in qualitative research,” International Journal of social research method-
ology, vol. 11, no. 4, pp. 327–344, 2008.
[27] M. Naderifar, H. Goli, and F. Ghaljaie, “Snowball sampling: A purpose-
ful method of sampling in qualitative research,” Strides in development
of medical education, vol. 14, no. 3, 2017.
[28] V. D. Blondel, J.-L. Guillaume, R. Lambiotte, and E. Lefebvre, “Fast
unfolding of communities in large networks,” Journal of statistical
mechanics: theory and experiment, vol. 2008, no. 10, p. P10008, 2008.
[29] G. S. Kido, R. A. Igawa, and S. Barbon Jr, “Topic modeling based on
louvain method in online social networks,” in Anais do XII Simp ´
osio
Brasileiro de Sistemas de Informac¸ ˜
ao. SBC, 2016, pp. 353–360.
[30] P. Joshi and S. Mohammed, “Identifying social media influencers using
graph analytics,” Authorea Preprints, 2023.
[31] T. Ganegedara, Natural Language Processing with TensorFlow: Teach
language to machines using Python’s deep learning library. Packt
Publishing Ltd, 2018.
[32] M. Koroteev, “Bert: a review of applications in natural language pro-
cessing and understanding,” arXiv preprint arXiv:2103.11943, 2021.
[33] L. Yue, W. Chen, X. Li, W. Zuo, and M. Yin, “A survey of sentiment
analysis in social media,” Knowledge and Information Systems, vol. 60,
pp. 617–663, 2019.
[34] Y. Li, M. A. Thomas, and D. Liu, “From semantics to pragmatics: where
is can lead in natural language processing (nlp) research,” European
Journal of Information Systems, vol. 30, no. 5, pp. 569–590, 2021.
[35] J. Gu, Z. Wang, J. Kuen, L. Ma, A. Shahroudy, B. Shuai, T. Liu,
X. Wang, G. Wang, J. Cai et al., “Recent advances in convolutional
neural networks,” Pattern recognition, vol. 77, pp. 354–377, 2018.
[36] H. Gholamalinezhad and H. Khosravi, “Pooling methods in deep neural
networks, a review,” arXiv preprint arXiv:2009.07485, 2020.
[37] J. T. Hancock and T. M. Khoshgoftaar, “Survey on categorical data for
neural networks,” Journal of Big Data, vol. 7, no. 1, pp. 1–41, 2020.
[38] Pontifical Council for Justice and Peace, Compendium of the Social
Doctrine of the Church. Vatican City: Libreria Editrice Vaticana, 2004,
sections 154 and 415.
[39] J. S. Mill, On Liberty. London: John W. Parker and Son, 1859, chapter
I and III.
[40] J. Rawls, A Theory of Justice, revised edition ed. Cambridge, MA:
Harvard University Press, 1999, see especially p. 3.
