Social media have become the ideal place for black hats and malicious individuals to target susceptible users through diferent attack vectors and then manipulate their opinions and interests. Fake news, radicalization, and pushing bias into the data represent some popular ways noxious users adopt to perpetrate their criminal intents. In this evolving scenario, Artificial Intelligence techniques represent a valuable tool to early detect and mitigate the risk due to the spreading of these emerging attacks. In this work, we describe the Machine Learning based solutions developed to address the problems mentioned above and our current research.
pability to efectively process diferent types of raw data (e.g., text, images, preferences, etc.) without the necessity Nowadays, the users’ opinions and interests can be ma- of a feature engineering phase and the intervention of nipulated through a wide range of attack vectors: spread- an expert. ing fake news, radicalization due to specific suggestions In this work, we provide an overview of the abovein recommender systems, and pushing biased informa- mentioned problems and discuss current and future retion in the training data for Machine Learning (ML) search lines. Specifically, in Section 2 we introduce the models represent just some recent examples of this phe- fake news detection problem and show two DL based nomenon. Monitoring and early detection of these mali- solutions; Section 3 describes the radicalization phecious behaviors are becoming crucial problems for politi- nomenon and presents a simulation framework to invescal organizations and institutions as they can manipulate tigate its efects; Section 4 depicts a solution to mitigate public opinion and change the results of events and cam- the risk due to presence of bias in data; finally, Section 5 paigns. concludes the work and propose some interesting new
In this complex and evolving scenario, Artificial Intel- research lines. ligence (AI) and Machine Learning techniques can play a key role in detecting and mitigating the risk due to these new emerging attacks. In particular, there is a growing interest in Deep Learning (DL) based solutions as their ca
of news represents a crucial problem that can benefit from recent advances in AI and ML. In particular, the automatic detection of fake news is a relevant problem attracting great interest from the research community. This problem was traditionally addressed in the literature as a text classification problem [ 5] i.e., distinguishing between real and fake news documents.
However, learning reliable detection models able to identify misinformation requires coping with diferent complex issues. First, an efective solution should allow for handling low-level raw data frequently afected by noise, since the channels used to spread fake news typically allow for sharing only short text (e.g., Twitter). Moreover, the number of labeled training instances is limited; the labeling phase is a dificult and time-consuming task manually performed by domain experts. Finally, fake news can concern diferent topics; therefore, the features leveraged to perform the prediction should be domain independent to handle diferent topics. In addition, malicious contents represent only a limited portion of the data; then, the training set will exhibit an unbalanced distribution that makes the learning phase of the model more dificult.
As mentioned above, in real scenarios data scarcity and the lack of labeled data can greatly afect the performances of Fake News Detectors. Recent approaches propose the usage of semi-supervised techniques to leverage the huge amounts of unlabelled data to boost the detection capability of the learned models. In [6], we proposed to employ a pre-trained instance of BERT model [7] as a backbone for classifying (as either fake or not) short news documents coming from a specific domain. However, instead of simply trying to fine-tune the BERT instance for this classification task, it is integrated into a self-training scheme. In more detail, a Pseudo-Labelling approach is used to map a number of unlabelled data instances, sampled from a given instance bucket. The pseudo-labels are assigned by a classification model, which is iteratively trained against a growing collection of both originallylabeled examples and pseudo-labeled ones. This process is repeated (in a self-training cycle) until a suitable stop criterion is satisfied (e.g., the maximum number of iterations, loss convergence, etc.).
This approach is depicted in Figure 1, where the initial set of labeled data is enriched with unlabeled data while using the classifier trained on the labeled data to assign “artificial” labels to unlabeled ones. Then, a new version of the classifier is built by using both the originally-labeled data and the newly automatically-labeled ones. Extensive experiments conducted on two diferent real datasets confirmed the efectiveness of the proposed approach in discovering accurate enough classification models even when the fraction of labeled data is relatively small.
Labeled Data
Unlabeled Data 1. Train the model
Classification
Model
New Classification
Model 3. Train a new model with pseudo labeled and labeled data 2. Predict labels for unlabeled data
Pseudo Labeled Data
detection is that they can vary across diferent domains.
The problem is particularly relevant as they may emerge in contexts for which no prior evidence is available [8, 9, 10, 11, 12, 13, 14]. Despite several proposals, state-ofthe-art exhibit several limitations. Most of the existing approaches are indeed designed for specific domains and, as a result, poorly perform in cross-domain scenarios.
To overcome this drawback, the proposed work introduces an end-to-end DL based framework for fake news detection tailored for cross-domain applications. The adoption of the DL paradigm [15] represents a natural solution to address the above issues, as it permits the learning of accurate classification models also from raw data (in our solution, the words composing the news) without requiring heavy intervention by data-science experts. Basically, these DL models are structured according to a hierarchical architecture (consisting of several layers of base computational units i.e., the artificial neurons are stacked one upon the other), allowing for learning features at diferent abstraction levels to represent raw data.
Relying on the success of deep neural networks, we combine recent advancements in the field, like autoencoders and adversarial generative approaches, for learning latent domain feature representations to use on new domains. Moreover, we derive features that are domaininvariant and, thus, benefit the detection of fake news on newly arrived, emergent events for which only a few verified news are available. ^ z z
Domain Discriminator
Embedding Reconstruction (auxiliary task)
The proposed solution is composed of three neural interactions in their preference history. Here, we investicomponents (see Figure 2) that collaborate to solve two gate whether and how RecSys can afect and modify the tasks simultaneously: the main one is to recognize fake users’ preferences, as they could be used to manipulate information, and the second (auxiliary) task aims to pro- their opinions and behaviors. In more detail, we devise duce domain invariant features. Preliminary experimen- a simulation framework for modeling the evolution of tation conducted on two real datasets shows promising the interactions between users and items. In our setting, results and encourages further studies. we assume that such interactions are fully driven by a (black-box) collaborative filtering algorithm. The simulation model is intended to start from an initial group 3. Radicalization in Recommender of heterogeneous user preferences, from which the recSystems ommender system induces initial transition probabilities between item categories. The interaction between the Today, digital platforms such as social media and e- simulation process and the recommendation algorithm commerce websites integrate advanced Recommender allows for simulating the evolution of such preferences. Systems (RecSys) to improve the user experience and By analyzing the resulting transition probabilities, we make the items search more efective. With the grow- are interested in assessing the impact of the algorithm ing amount of available content, RecSys have become on the initial population of user preferences. a valuable tool to support users in navigating the large In a nutshell, given the initial users interactions, our volumes of information proposed by these platforms. goal is to generate a probabilistic graph , whose nodes
Alas, the social consequences deriving from their us- include the subset of items for which the user had at age are still object of study. Some recent works analyze least one interaction during the simulation. their impact in terms of harmful phenomena such as In this way, estimating the users’ leaning drift built on echo chambers [16, 17], rabbit holes [18, 19], filter bub- this graph corresponds to estimating the recommender’s bles [20], and radicalization [21, 22, 23], showing that influence in the long term. Figure 3 depicts an overview the long-term interaction with the RecSys can drift users of the simulation framework we intend to implement. towards polarized opinions and noxious contents. As Our final aim is to evaluate how much the recommena relevant example, it has been observed that continu- dation algorithm changes user preferences after a certain ous interactions with the recommendation algorithm can number of interactions. We refer to this tendency as lead to severe extremization of users’ political leaning algorithmic drift. [24]. In practice, given a recommendation algorithm pre
In the following, we describe our current research line trained with the initial user interactions (classified as consisting of a simulation framework to evaluate the neutral or harmful), the RecSys will tend to suggest more efects of RecSys on the users’ preferences. extreme content to initially unbiased users, and, viceversa, neutral and harmless content to already radicalized
To quantify the algorithmic drift induced by the recomIn our research, we consider a typical recommendation mender system, we intend to define novel graph-based scenario in which some top-ranked items [25] are pro- metrics to be applied over the probabilistic graph posed and accepted by the user [26]. Items (that can of a user . Basically, we are interested in measuring take the form of media, news, videos, etc.) are tagged as the probability to move from harmful to neutral items harmful or neutral, while users are categorized as non-, (resp. from neutral to harmful) and to remain in cliques semi- or radicalized, based on the percentage of harmful of neutral (resp. harmful) ones. As a consequence, we
want to estimate the change in users’ leaning before and regulate the re-ranking process, we introduced a hyperafter interacting with the recommendations. parameter to tune the importance of the re-ranking. By
The underlying idea is that the larger the value, the applying this mitigation strategy, we expect to reduce the more the recommender is shifting the user’s choices to- noxious exposure obtained in the long term, especially wards harmful items with the respect to its initial prefer- towards initially unbiased users. Moreover, unlike the ences. pre-processing strategy, this one exhibits the relevant benefit that does not require additional retraining phases of the model.
basis of diferent strategies. Here, we define two diferent approaches: a pre-processing mitigation strategy and a post-processing one.
Pre-processing based solution. This strategy aims to modify the initial interactions dataset prior to training the recommender. As it can be viewed as a bipartite graph consisting of connections between users and items, we hypothesize that hindering the flow between nonradicalized users and items on the bipartite graph will also hinder the noxious efects underlying algorithmic drift of non-radicalized users in the long term.
This approach is implemented by adopting a rewiring strategy that, for each non-radicalized user , randomly deletes a percentage % of its edges toward items shared with semi-radicalized users. It then rewires them randomly choosing from the subset of neutral items which are exclusively connected to other non-radicalized users, i.e., an edge does not exist between these neutral items and semi-radicalized or radicalized users.
Once the rewiring is done for all non-radicalized users, we obtain a new altered dataset, which we use for training the recommender.
Post-processing based strategy. Based on the solution proposed in [27] for popularity debiasing, we propose to penalize the recommendation scores assigned to items. The rationale lies in boosting the final recommendation scores associated with neutral items and reducing the ones associated with harmful items. In practice, for each user, the penalization assigned to harmful items in the final recommendation score is inversely related to the radicalization level of the user (i.e., the percentage of harmful items in their initial history). To more accurately
a recent research trend focuses on balancing the quality performances with ethical or discriminatory efects of ML-based recommendations. Indeed, unfairness in Decision Support Systems and RecSys allows for performing sophisticated attacks able to manipulate users’ opinions or segregate subpopulations based on irrelevant characteristics [28]. Typical examples are limited access to credit or healthcare due, e.g., to race or gender. For example, a recent research1 highlighted that an algorithm adopted to predict patients likely to need extra medical care heavily favored white patients over black patients.
Unfairness may come into play mainly for two reasons: inductive biases of algorithms (i.e., inheriting intrinsic biases of data) or through adversarial attacks (i.e., when a malicious actor purposefully poisons an ML system by exploiting some model weakness).
In [29], we addressed this problem in the case of RecSys. Specifically, we modeled a DL architecture that significantly improves the exposure of low-popular items. The proposed technique is based on two main aspects: resampling negative items and ensembling multiple instances of the algorithm.
ICAR-CNR is developing in the framework of the SERICS project to mitigate the risk due to attacks aiming at manipulating the users’ behaviors. Specifically, we have
a-major-health-care-risk-algorithm/. shown as AI-based tools could represent a promising tool [8] K. Shu, L. Cui, S. Wang, D. Lee, H. Liu, Defend: to detect, evaluate and reduce the influence of these new Explainable fake news detection, in: Proceedings emerging attacks. of the 25th ACM SIGKDD International Conference
As future works, we are interested in extending our on Knowledge Discovery and Data Mining, KDD preliminary studies concerning radicalization phenom- ’19, 2019, p. 395–405. ena, relying on more sophisticated and realistic simula- [9] C. Raj, P. Meel, Arcnn framework for multimodal tion frameworks. In addition, as regards the fake news infodemic detection, Neural Networks 146 (2022) detection task, we aim to exploit multi-modal data like 36–68. images, video, news propagation patterns, social context [10] T. Sachan, N. Pinnaparaju, M. Gupta, V. Varma, representing the user engagements of news on social me- Scate: Shared cross attention transformer encoders dia (e.g. the number of followers, hashtags, friendship for multimodal fake news detection, in: Proceednetworks, retweets) to improve the detection capability ings of the 2021 IEEE/ACM International Conferof the neural models. ence on Advances in Social Networks Analysis and Mining, ASONAM ’21, 2021, p. 399–406. [11] R. Kumari, A. Ekbal, Amfb: Attention based mulAcknowledgments timodal factorized bilinear pooling for multimodal fake news detection, Expert Systems with ApplicaThis work was partially supported by project SERICS tions 184 (2021) 115412. (PE00000014) under the NRRP MUR program funded by [12] Z. Jin, J. Cao, H. Guo, Y. Zhang, J. Luo, Multimodal the EU - NGEU. fusion with recurrent neural networks for rumor detection on microblogs, in: Proceedings of the References 25th ACM International Conference on Multimedia, Association for Computing Machinery, New York, [1] X. Zhou, R. Zafarani, A survey of fake news: Fun- NY, USA, 2017, p. MM ’17.
damental theories, detection methods, and oppor- [13] Q. Jing, D. Yao, X. Fan, B. Wang, H. Tan, X. Bu, J. Bi, tunities, ACM Comput. Surv. 53 (2020). Transfake: Multi-task transformer for multimodal [2] H. Allcott, M. Gentzkow, Social media and fake enhanced fake news detection, in: IJCNN, 2021, pp. news in the 2016 election, Journal of Economic 1–8.
Perspectives 31 (2017) 211–36. [14] J. Wang, H. Mao, H. Li, Fmfn: Fine-grained mul[3] H. MARSHALL, A. DRIESCHOVA, Post-truth poli- timodal fusion networks for fake news detection, tics in the uk’s brexit referendum, New Perspectives Applied Sciences 12 (2022).
26 (2018) 89–106. [15] Y. Le Cun, Y. Bengio, G. Hinton, Deep learning, [4] C. Ngai, R. Singh, L. Yao, Impact of covid-19 vaccine Nature 521 (2015) 436–444.
misinformation on virality on social media: Content [16] E. Pariser, The filter bubble: How the new personanalysis of message themes and writing strategies alized web is changing what we read and how we (preprint), Journal of Medical Internet Research 24 think, Penguin, 2011.
(2022). [17] H. Ferraz de Arruda, F. Maciel Cardoso, G. Ferraz [5] C. Liu, X. Wu, M. Yu, G. Li, J. Jiang, W. Huang, de Arruda, A. R. Hernández, L. da Fontoura Costa, X. Lu, A two-stage model based on bert for short Y. Moreno, Modelling how social network algofake news detection, in: C. Douligeris, D. Kara- rithms can influence opinion polarization, Informagiannis, D. Apostolou (Eds.), Knowledge Science, tion Sciences 588 (2022) 265–278.
Engineering and Management, Springer Interna- [18] M. H. Ribeiro, R. Ottoni, R. West, V. A. Almeida, tional Publishing, Cham, 2019, pp. 172–183. W. Meira Jr, Auditing radicalization pathways on [6] P. Zicari., M. Guarascio., L. Pontieri., G. Folino., youtube, in: Proceedings of the 2020 conference on Learning deep fake-news detectors from scarcely- fairness, accountability, and transparency, 2020, pp. labelled news corpora, in: Proceedings of 131–141. the 25th International Conference on Enterprise [19] M. Haroon, A. Chhabra, X. Liu, P. Mohapatra, Information Systems - Volume 1: ICEIS„ IN- Z. Shafiq, M. Wojcieszak, Youtube, the great radSTICC, SciTePress, 2023, pp. 344–353. doi:10. icalizer? auditing and mitigating ideological bi5220/0011827500003467. ases in youtube recommendations, arXiv preprint [7] J. Devlin, M. W. Chang, K. Lee, K. Toutanova, BERT: arXiv:2203.10666 (2022).
Pre-training of deep bidirectional transformers for [20] H. Zhang, Z. Zhu, J. Caverlee, Evolution of filter language understanding, in: NAACL-HLT, Associa- bubbles and polarization in news recommendation, tion for Computational Linguistics, 2019, pp. 4171– 2023. arXiv:2301.10926. 4186. [21] H. Hosseinmardi, A. Ghasemian, A. Clauset, M. Mobius, D. M. Rothschild, D. J. Watts, Examining the rank, Association for Computing Machinery, New consumption of radical content on youtube, Pro- York, NY, USA, 2019. ceedings of the National Academy of Sciences 118 [27] H. Abdollahpouri, Popularity bias in ranking and (2021) e2101967118. recommendation, AIES ’19: Proceedings of the 2019 [22] M. H. Ribeiro, V. Veselovsky, R. West, The ampli- AAAI/ACM Conference on AI, Ethics, and Society ifcation paradox in recommender systems, 2023. (2019) 529–530.
arXiv:2302.11225. [28] L. Oneto, N. Navarin, B. Biggio, F. Errica, A. Micheli, [23] F. Fabbri, Y. Wang, F. Bonchi, C. Castillo, M. Math- F. Scarselli, M. Bianchini, L. Demetrio, P. Bongini, ioudakis, Rewiring what-to-watch-next recommen- A. Tacchella, A. Sperduti, Towards learning trustdations to reduce radicalization pathways, Asso- worthily, automatically, and with guarantees on ciation for Computing Machinery, New York, NY, graphs: An overview, Neurocomputing 493 (2022) USA, 2022. 217–243. [24] M. Haroon, A. Chhabra, X. Liu, P. Mohapatra, [29] L. Caroprese, G. Manco, M. Minici, F. S. Pisani, Z. Shafiq, M. Wojcieszak, Youtube, the great radi- E. Ritacco, Unbiasing collaborative filtering for calizer? auditing and mitigating ideological biases popularity-aware recommendation (discussion pain youtube recommendations (2022). per), in: Proceedings of the 29th Italian Symposium [25] Z. Guan, E. Cutrell, An eye tracking study of the on Advanced Database Systems, SEBD 2021, Pizzo efect of target rank on web search, Association for Calabro (VV), Italy, September 5-9, 2021, volume Computing Machinery, New York, NY, USA, 2007. 2994 of CEUR Workshop Proceedings, CEUR-WS.org, [26] A. Agarwal, X. Wang, C. Li, M. Bendersky, M. Na- 2021, pp. 450–457.
jork, Addressing trust bias for unbiased learning-to