This article examines 3,517 Facebook ads created by Russia’s Internet Research Agency (IRA) between June 2015 and August 2017 in its Active Measures disinformation campaign targeting the 2016 U.S. presidential election. We aimed to unearth the relationship between ad engagement (ad clicks) and 40 features related to the ads’ metadata, psychological meaning, and sentiment. The purpose of our analysis was to (1) understand the relationship between engagement and features, (2) find the most relevant feature subsets to predict engagement via feature selection, and (3) find the semantic topics that best characterize the data set via topic modeling. We found that investment features (e.g., ad spend, ad lifetime), caption length, and sentiment were the top features predicting users’ engagement with the ads. In addition, positive sentiment ads were more engaging than negative ads, and psycholinguistic features (e.g., use of religion-relevant words) were identified as highly important in the makeup of an engaging disinformation ad. Linear support vector machines (SVMs) and logistic regression classifiers achieved the highest mean F scores (93.6%), revealing that the optimal feature subset contains 12 and six features, respectively. Finally, we corroborate the findings of previous research that the IRA specifically targeted Americans on divisive ad topics (e.g., LGBT rights) and advance a definition of disinformation advertising.

Misinformation entails the dissemination of falsehoods that leads to the slow fracturing of society via decreased trust in democratic processes, institutions, and science. The public has grown aware of the role of social media as a superspreader of untrustworthy information, where even pandemics have not been immune. In this paper, we focus on COVID-19 misinformation and examine a subset of 2.1M tweets to understand misinformation as a function of engagement, tweet content (COVID-19- vs. non-COVID-19-related), and veracity (misleading or factual). Using correlation analysis, we show the most relevant feature subsets among over 126 features that most heavily correlate with misinformation or facts. We found that (i) factual tweets, regardless of whether COVID-related, were more engaging than misinformation tweets; and (ii) features that most heavily correlated with engagement varied depending on the veracity and content of the tweet.

This paper investigates the use of machine learning models to classify unhealthy online conversations containing one or more forms of subtler abuse, such as hostility, sarcasm, and generalization. We leveraged a public dataset of 44K online comments containing healthy and unhealthy comments labeled with seven forms of subtle toxicity. We were able to distinguish between these comments with a micro F1-score, macro F1-score, and ROC-AUC of 88.76%, 67.98%, and 0.71, respectively. Hostile comments were easier to detect than other types of unhealthy comments. We also conducted a sentiment analysis that revealed that most unhealthy comments were associated with a slight negative sentiment, with hostile comments being the most negative.

This paper investigates whether computer usage profiles comprised of process-, network-, mouse-, and keystroke-related events are unique and consistent over time in a naturalistic setting, discussing challenges and opportunities of using such profiles in applications of continuous authentication. We collected ecologically-valid computer usage profiles from 31 MS Windows 10 computer users over 8 weeks and submitted this data to comprehensive machine learning analysis involving a diverse set of online and offline classifiers. We found that: (i) profiles were mostly consistent over the 8-week data collection period, with most (83.9%) repeating computer usage habits on a daily basis; (ii) computer usage profiling has the potential to uniquely characterize computer users (with a maximum F-score of 99.90%); (iii) network-related events were the most relevant features to accurately recognize profiles (95.69% of the top features distinguishing users were network-related); and (iv) binary models were the most well-suited for profile recognition, with better results achieved in the online setting compared to the offline setting (maximum F-score of 99.90% vs. 95.50%).

Phishing and disinformation are popular social engineering attacks with attackers invariably applying influence cues in texts to make them more appealing to users. We introduce Lumen, a learning-based framework that exposes influence cues in text: (i) persuasion, (ii) framing, (iii) emotion, (iv) objectivity/subjectivity, (v) guilt/blame, and (vi) use of emphasis. Lumen was trained with a newly developed dataset of 3K texts comprised of disinformation, phishing, hyperpartisan news, and mainstream news. Evaluation of Lumen in comparison to other learning models showed that Lumen and LSTM presented the best F1-micro score, but Lumen yielded better interpretability. Our results highlight the promise of ML to expose influence cues in text, toward the goal of application in automatic labeling tools to improve the accuracy of human-based detection and reduce the likelihood of users falling for deceptive online content.

Open-source software (OSS) is a critical part of the software supply chain. Recent social engineering attacks against OSS development teams have enabled attackers to become code contributors and later inject malicious code or vulnerabilities into the project with the goal of compromising dependent software. The attackers have exploited interactions among development team members and the social dynamics of team behavior to enable their attacks. We introduce a security approach that leverages signatures and patterns of team dynamics to predict the susceptibility of a software development team to social engineering attacks that enable access to the OSS project code. The proposed approach is programming language-, platform-, and vulnerability-agnostic because it assesses the artifacts of OSS team interactions, rather than OSS code.

This paper investigates whether computer usage profiles comprised of process-, network-, mouse- and keystroke-related events are unique and temporally consistent in a naturalistic setting, discussing challenges and opportunities of using such profiles in applications of continuous authentication. We collected ecologically-valid computer usage profiles from 28 MS Windows 10 computer users over 8 weeks and submitted this data to comprehensive machine learning analysis involving a diverse set of online and offline classifiers. We found that (i) computer usage profiles have the potential to uniquely characterize computer users (with a maximum F-score of 99.94%); (ii) network-related events were the most useful features to properly recognize profiles (95.14% of the top features distinguishing users being network-related); (iii) user profiles were mostly inconsistent over the 8-week data collection period, with 92.86% of users exhibiting drifts in terms of time and usage habits; and (iv) online models are better suited to handle computer usage profiles compared to offline models (maximum F-score for each approach was 95.99% and 99.94%, respectively).

This paper examines 3,517 Facebook ads created by Russia's Internet Research Agency (IRA) between June 2015 and August 2017 in its active measures disinformation campaign targeting the 2016 U.S. general election. We aimed to unearth the relationship between ad engagement (as measured by ad clicks) and 41 features related to ads' metadata, sociolinguistic structures, and sentiment. Our analysis was three-fold: (i) understand the relationship between engagement and features via correlation analysis; (ii) find the most relevant feature subsets to predict engagement via feature selection; and (iii) find the semantic topics that best characterize the dataset via topic modeling. We found that ad expenditure, text size, ad lifetime, and sentiment were the top features predicting users' engagement to the ads. Additionally, positive sentiment ads were more engaging than negative ads, and sociolinguistic features (e.g., use of religion-relevant words) were identified as highly important in the makeup of an engaging ad. Linear SVM and Logistic Regression classifiers achieved the highest mean F-scores (93.6% for both models), determining that the optimal feature subset contains 12 and 6 features, respectively. Finally, we corroborate the findings of related works that the IRA specifically targeted Americans on divisive ad topics (e.g., LGBT rights, African American reparations).

Disinformation entails the purposeful dissemination of falsehoods towards a greater dubious agenda and the chaotic fracturing of a society. The general public has grown aware of the misuse of social media towards these nefarious ends, where even global public health crises have not been immune to misinformation (deceptive content spread without intended malice). In this paper, we examine nearly 505K COVID-19-related tweets from the initial months of the pandemic to understand misinformation as a function of bot-behavior and engagement. Using a correlation-based feature selection method, we selected the 11 most relevant feature subsets among over 170 features to distinguish misinformation from facts, and to predict highly engaging misinformation tweets about COVID-19. We achieved an average F-score of at least 72\% with ten popular multi-class classifiers, reinforcing the relevance of the selected features. We found that (i) real users tweet both facts and misinformation, while bots tweet proportionally more misinformation; (ii) misinformation tweets were less engaging than facts; (iii) the textual content of a tweet was the most important to distinguish fact from misinformation while (iv) user account metadata and human-like activity were most important to predict high engagement in factual and misinformation tweets; and (v) sentiment features were not relevant.

The Microsoft Kinect® sensor has been employed for developing serious games and for biomechanics analysis. Both applications, when combined in the context of motor rehabilitation, might provide relevant data for therapists. However, the reliability of clinical data obtained with Kinect® is affected by filtering parameters which should be chosen according to spectral characteristics of the signals. In this paper we aim at determining the spectral characteristics of kinematics data collected with Kinect® during a serious game and to suggest adequate filtering. The motor tasks of lateral trunk inclination, trunk rotation, and shoulder abduction performed with heading, ski, and goalkeeper games originated 45 time series derived from 5 healthy people and 87 time series of 4 people with stroke. Time series were analyzed using the Fourier analysis and empirical mode decomposition (EMD). A residual analysis was performed to determine the optimal cutoff frequencies of the fourth-order low-pass Butterworth filters. Fourier and EMD analyses evidenced that the highest spectral power for header and goalkeeper tasks is below 3 Hz and for skiing, it is below 0.8 Hz. The ideal cutoff frequencies were around 3 Hz and 5 Hz and differed between healthy and stroke groups. The range of motion was affected by the cutoff frequencies. The signals captured by Kinect® have the main spectral components at lower frequencies and should be filtered at cutoff frequencies below 6 Hz. We recommend including the determination the impact of signal processing on clinical indicators in the workflow when developing a serious game for rehabilitation.

This study examines the influence of time series duration on the discriminative power of center-of-pressure (COP) features in distinguishing different population groups via statistical tests and machine learning (ML) models. We used two COP datasets, each containing two groups. One was collected from older adults with low or high risk of falling (dataset I), and the other from healthy and post-stroke adults (dataset II). Each time series was mapped into a vector of 34 features twice: firstly, using the original duration of 60 s, and then using only the first 30 s. We then compared each feature across groups through traditional statistical tests. Next, we trained six popular ML models to distinguish between the groups using features from the original signals and then from the shorter signals. The performance of each ML model was then compared across groups for the 30 s and 60 s time series. The mean percentage of features able to discriminate the groups via statistical tests was 26.5% smaller for 60 s signals in dataset I, but 13.5% greater in dataset II. In terms of ML, better performances were achieved for signals of 60 s in both datasets, mainly for similarity-based algorithms. Hence, we recommend the use of COP time series recorded over at least 60 s. The contribution of this paper also include insights into the robustness of popular ML models to the sampling duration of COP time series.

The postural control can be investigated from time series data of the center-of-pressure (COP) displacements. Detrended fluctuation analysis and scaled windowed variance are commonly employed to measure fractality in COP signal, while sample entropy and multiscale sample entropy are often used to address its regularity and complexity, respectively. Based on COP data from 19 post stroke adults and 19 healthy matched subjects, we first support previous findings that the sampling and/or the digital filtering of those data may influence the interpretations on postural control provided by such types of metrics. Then, we show evidences that the digital filtering lead to less accurate information on the entropy in postural sway with either traditional statistical tools or the decision tree (DT) classifier. Thus, when computing entropy-related features, it is not advisable to filter the data. However, if fractal features are considered instead, the use of digital filters and downsampling techniques can provide more discriminative information. When combining fractal and entropy-related features, both original and processed COP data should be considered for either DT or other popular classifiers. Lastly, with the aid of a DT, we could classify the individuals with an accuracy of 77.6% for fractal features only (best case), 68.4% for entropy-related features only, and 76.3% after combining fractal and entropy-related features.

Maintaining balance is a motor task of crucial importance for humans to perform their daily activities safely and independently. Studies in the field of Artificial Intelligence have considered different classification methods in order to distinguish healthy subjects from patients with certain motor disorders based on their postural strategies during the balance control. The main purpose of this paper is to compare the performance between Decision Tree (DT) and Genetic Programming (GP) -both classification methods of easy interpretation by health professionals - to distinguish postural sway patterns produced by healthy and stroke individuals based on 16 widely used posturographic variables. For this purpose, we used a posturographic dataset of time-series of center-of-pressure displacements derived from 19 stroke patients and 19 healthy matched subjects in three quiet standing tasks of balance control. Then, DT and GP models were trained and tested under two different experiments where accuracy, sensitivity and specificity were adopted as performance metrics. The DT method has performed statistically significant (P < 0.05) better in both cases, showing for example an accuracy of 72.8% against 69.2% from GP in the second experiment of this paper.

Since balance control is a basic prerequisite for most of our daily activities, this task has crucial importance in the functional independence of humans. During the balance control, the human body sways constantly, even in the quiet upright stance. This body sway is usually captured in the form of time series of center-of-pressure (COP) displacements with the help of a measurement device known as force platform. In this paper, rather than using the traditional statistical analysis widely found in balance assessment studies, machine learning techniques were employed to recognize stroke patients and healthy matched subjects based on posturographic features extracted from their COP data. In this context, our main purpose was to investigate the relevance of 16 linear and 9 nonlinear posturographic features commonly examined in the balance assessment field. Thus, the average joint performance among six popular classification methods was evaluated under a 65 instances-size dataset in three situations: using only linear features, only nonlinear features and, finally, linear and nonlinear posturographic features combined. The former situation yielded significantly (P <;0.01) better results. This finding suggest that, following an approach based on classification methods to distinguish healthy from stroke physiological systems, the overall amount of sway indexed by the linear features is more relevant than the temporal patterns of sway described by the nonlinear features.