arXiv subjects

Nikos Salamanos

Publications and source records attributed to Nikos Salamanos.

7 recordsLinked to original sources

arXiv2026

DeTox-Fed: Detecting Toxic Conversations in the Fediverse with Federated Graph Neural Networks

The rise of decentralized social networks (DSNs), and in particular the rapid uptake of the Fediverse (e.g., Pleroma, Mastodon, Lemygrad), introduces new challenges in content moderation. Independent instances host their own data, follow different moderation policies, and often observe only partial views of conversations. We present DeTox-Fed, a federated graph-learning framework for detecting toxic conversations in DSNs without requiring instances to share raw conversations or moderation labels. Each instance constructs a local conversation graph, where nodes represent conversation trees and edges capture shared user participation across conversations. A Graph Neural Network (GNN) is then trained in a federated learning setup, allowing instances to collaboratively learn a toxicity classifier while preserving data locality. Unlike text-only moderation approaches, DeTox-Fed combines conversational structure, user-interaction patterns, conversation-level statistics, and aggregate sentiment signals. We evaluate the framework on a large Pleroma conversation dataset and show that it achieves stable toxic conversation detection under limited local labels, partial client participation, and varying moderation thresholds. Our results indicate that federated graph-based moderation is a promising direction for semi-automated moderation in decentralized social networks.

cs.SI↗

arXiv2024

HyperGraphDis: Leveraging Hypergraphs for Contextual and Social-Based Disinformation Detection

In light of the growing impact of disinformation on social, economic, and political landscapes, accurate and efficient identification methods are increasingly critical. This paper introduces HyperGraphDis, a novel approach for detecting disinformation on Twitter that employs a hypergraph-based representation to capture (i) the intricate social structures arising from retweet cascades, (ii) relational features among users, and (iii) semantic and topical nuances. Evaluated on four Twitter datasets -- focusing on the 2016 U.S. Presidential election and the COVID-19 pandemic -- HyperGraphDis outperforms existing methods in both accuracy and computational efficiency, underscoring its effectiveness and scalability for tackling the challenges posed by disinformation dissemination. HyperGraphDis displays exceptional performance on a COVID-19-related dataset, achieving an impressive F1 score (weighted) of approximately 89.5%. This result represents a notable improvement of around 4% compared to the other state-of-the-art methods. Additionally, significant enhancements in computation time are observed for both model training and inference. In terms of model training, completion times are accelerated by a factor ranging from 2.3 to 7.6 compared to the second-best method across the four datasets. Similarly, during inference, computation times are 1.3 to 6.8 times faster than the state-of-the-art.

cs.SI↗

arXiv2023

Identifying Misinformation on YouTube through Transcript Contextual Analysis with Transformer Models

Misinformation on YouTube is a significant concern, necessitating robust detection strategies. In this paper, we introduce a novel methodology for video classification, focusing on the veracity of the content. We convert the conventional video classification task into a text classification task by leveraging the textual content derived from the video transcripts. We employ advanced machine learning techniques like transfer learning to solve the classification challenge. Our approach incorporates two forms of transfer learning: (a) fine-tuning base transformer models such as BERT, RoBERTa, and ELECTRA, and (b) few-shot learning using sentence-transformers MPNet and RoBERTa-large. We apply the trained models to three datasets: (a) YouTube Vaccine-misinformation related videos, (b) YouTube Pseudoscience videos, and (c) Fake-News dataset (a collection of articles). Including the Fake-News dataset extended the evaluation of our approach beyond YouTube videos. Using these datasets, we evaluated the models distinguishing valid information from misinformation. The fine-tuned models yielded Matthews Correlation Coefficient>0.81, accuracy>0.90, and F1 score>0.90 in two of three datasets. Interestingly, the few-shot models outperformed the fine-tuned ones by 20% in both Accuracy and F1 score for the YouTube Pseudoscience dataset, highlighting the potential utility of this approach -- especially in the context of limited training data.

cs.CL↗

arXiv2022

Privacy-Preserving Online Content Moderation: A Federated Learning Use Case

Users are daily exposed to a large volume of harmful content on various social network platforms. One solution is developing online moderation tools using Machine Learning techniques. However, the processing of user data by online platforms requires compliance with privacy policies. Federated Learning (FL) is an ML paradigm where the training is performed locally on the users' devices. Although the FL framework complies, in theory, with the GDPR policies, privacy leaks can still occur. For instance, an attacker accessing the final trained model can successfully perform unwanted inference of the data belonging to the users who participated in the training process. In this paper, we propose a privacy-preserving FL framework for online content moderation that incorporates Differential Privacy (DP). To demonstrate the feasibility of our approach, we focus on detecting harmful content on Twitter - but the overall concept can be generalized to other types of misbehavior. We simulate a text classifier - in FL fashion - which can detect tweets with harmful content. We show that the performance of the proposed FL framework can be close to the centralized approach - for both the DP and non-DP FL versions. Moreover, it has a high performance even if a small number of clients (each with a small number of data points) are available for the FL training. When reducing the number of clients (from 50 to 10) or the data points per client (from 1K to 0.1K), the classifier can still achieve ~81% AUC. Furthermore, we extend the evaluation to four other Twitter datasets that capture different types of user misbehavior and still obtain a promising performance (61% - 80% AUC). Finally, we explore the overhead on the users' devices during the FL training phase and show that the local training does not introduce excessive CPU utilization and memory consumption overhead.

cs.LG↗

arXiv2021

A Unified Graph-Based Approach to Disinformation Detection using Contextual and Semantic Relations

As recent events have demonstrated, disinformation spread through social networks can have dire political, economic and social consequences. Detecting disinformation must inevitably rely on the structure of the network, on users particularities and on event occurrence patterns. We present a graph data structure, which we denote as a meta-graph, that combines underlying users' relational event information, as well as semantic and topical modeling. We detail the construction of an example meta-graph using Twitter data covering the 2016 US election campaign and then compare the detection of disinformation at cascade level, using well-known graph neural network algorithms, to the same algorithms applied on the meta-graph nodes. The comparison shows a consistent 3%-4% improvement in accuracy when using the meta-graph, over all considered algorithms, compared to basic cascade classification, and a further 1% increase when topic modeling and sentiment analysis are considered. We carry out the same experiment on two other datasets, HealthRelease and HealthStory, part of the FakeHealth dataset repository, with consistent results. Finally, we discuss further advantages of our approach, such as the ability to augment the graph structure using external data sources, the ease with which multiple meta-graphs can be combined as well as a comparison of our method to other graph-based disinformation detection frameworks.

cs.SI↗

arXiv2021

Did State-sponsored Trolls Shape the 2016 US Presidential Election Discourse? Quantifying Influence on Twitter

It is a widely accepted fact that state-sponsored Twitter accounts operated during the 2016 US presidential election, spreading millions of tweets with misinformation and inflammatory political content. Whether these social media campaigns of the so-called "troll" accounts were able to manipulate public opinion is still in question. Here, we quantify the influence of troll accounts on Twitter by analyzing 152.5 million tweets (by 9.9 million users) from that period. The data contain original tweets from 822 troll accounts identified as such by Twitter itself. We construct and analyse a very large interaction graph of 9.3 million nodes and 169.9 million edges using graph analysis techniques, along with a game-theoretic centrality measure. Then, we quantify the influence of all Twitter accounts on the overall information exchange as is defined by the retweet cascades. We provide a global influence ranking of all Twitter accounts and we find that one troll account appears in the top-100 and four in the top-1000. This combined with other findings presented in this paper constitute evidence that the driving force of virality and influence in the network came from regular users - users who have not been classified as trolls by Twitter. On the other hand, we find that on average, troll accounts were tens of times more influential than regular users were. Moreover, 23% and 22% of regular accounts in the top-100 and top-1000 respectively, have now been suspended by Twitter. This raises questions about their authenticity and practices during the 2016 US presidential election.

cs.SI↗

arXiv2019

On the Influence of Twitter Trolls during the 2016 US Presidential Election

It is a widely accepted fact that state-sponsored Twitter accounts operated during the 2016 US presidential election spreading millions of tweets with misinformation and inflammatory political content. Whether these social media campaigns of the so-called "troll" accounts were able to manipulate public opinion is still in question. Here we aim to quantify the influence of troll accounts and the impact they had on Twitter by analyzing 152.5 million tweets from 9.9 million users, including 822 troll accounts. The data collected during the US election campaign, contain original troll tweets before they were deleted by Twitter. From these data, we constructed a very large interaction graph; a directed graph of 9.3 million nodes and 169.9 million edges. Recently, Twitter released datasets on the misinformation campaigns of 8,275 state-sponsored accounts linked to Russia, Iran and Venezuela as part of the investigation on the foreign interference in the 2016 US election. These data serve as ground-truth identifier of troll users in our dataset. Using graph analysis techniques we qualify the diffusion cascades of web and media context that have been shared by the troll accounts. We present strong evidence that authentic users were the source of the viral cascades. Although the trolls were participating in the viral cascades, they did not have a leading role in them and only four troll accounts were truly influential.

cs.SI↗