Authentication and de-authentication phases should occur at the beginning and end of secure user sessions, respectively. A secure session requires the user to pass the former, but the latter is often underestimated or ignored. Unattended or dangling sessions expose users to well-known Lunchtime Attacks. To mitigate this threat, researchers focused on automated de-authentication systems, either as a stand-alone mechanism or as a result of continuous authentication failures. Unfortunately, no single approach offers security, privacy, and usability. Face-recognition methods, for example, may be suitable for security and usability, but they violate user privacy by continuously recording their actions and surroundings. In this work, we propose BLUFADER, a novel continuous authentication system that takes advantage of blurred face detection and recognition to fast, secure, and transparent de-authenticate users, preserving their privacy. We obfuscate a webcam with a physical blur layer and use deep learning algorithms to perform face detection and recognition continuously. To evaluate BLUFADER's practicality, we collected two datasets formed by 30 recruited subjects (users) and thousands of physically blurred celebrity photos. The de-authentication system was trained and evaluated using the former, while the latter was used to appraise the privacy and increase variance at training time. To guarantee the privacy-preserving effectiveness of the selected physical blurring filter, we show that state-of-the-art deblurring models are not able to revert our physical blur. Further, we demonstrate that our approach outperforms state-of-the-art methods in detecting blurred faces, achieving up to 95% accuracy. Moreover, BLUFADER effectively de-authenticates users up to 100% accuracy in under 3 seconds, while satisfying security, privacy, and usability requirements. Last, our continuous authentication face recognition module based on Siamese Neural Network preventively protect users from adversarial attacks, enhancing the overall system security.

Ideally, secure user sessions should start and end with authentication and de-Authentication phases, respectively. While the user must pass the former to start a secure session, the latter's importance is often ignored or underestimated. Dangling or unattended sessions expose users to well-known Lunchtime Attacks. To mitigate this threat, the research community focused on automated de-Authentication systems. Unfortunately, no single approach offers security, privacy, and usability. For instance, although facial recognition-based methods might be a good fit for security and usability, they violate user privacy by constantly recording the user and the surrounding environment.In this work, we propose BLUFADE, a fast, secure, and transparent de-Authentication system that takes advantage of blurred faces to preserve user privacy. We obfuscate a webcam with a physical blur layer and use deep learning algorithms to perform face detection continuously. To assess BLUFADE's practicality, we collected two datasets formed by 30 recruited subjects (users) and thousands of physically blurred celebrity photos. The former was used to train and evaluate the deauthentication system performances, the latter to assess the privacy and to increase variance in training data. We show that our approach outperforms state-of-The-Art methods in detecting blurred faces, achieving up to 95% accuracy. Furthermore, we demonstrate that BLUFADE effectively de-Authenticates users up to 100% accuracy in under 3 seconds, while satisfying security, privacy, and usability requirements.

Voice messages are an increasingly popular method of communication, accounting for more than 200 million messages a day. Sending audio messages requires a user to invest lesser effort than texting while enhancing the message’s meaning by adding an emotional context (e.g., irony). Unfortunately, we suspect that voice messages might provide much more information than intended to prying ears of a listener. In fact, speech audio waves are both directly recorded by the microphone and propagated into the environment, and possibly reflected back to the microphone. Reflected waves along with ambient noise are also recorded by the microphone and sent as part of the voice message. In this paper, we propose a novel attack for inferring detailed information about user location (e.g., a specific room) leveraging a simple WhatsApp voice message. We demonstrated our attack considering 7,200 voice messages from 15 different users and four environments (i.e., three bedrooms and a terrace). We considered three realistic attack scenarios depending on previous knowledge of the attacker about the victim and the environment. Our thorough experimental results demonstrate the feasibility and efficacy of our proposed attack. We can infer the location of the user among a pool of four known environments with 85% accuracy. Moreover, our approach reaches an average accuracy of 93% in discerning between two rooms of similar size and furniture (i.e., two bedrooms) and an accuracy of up to 99% in classifying indoor and outdoor environments.

Personal Identification Numbers (PINs) are the most common user authentication method for in-person banking transactions at ATMs. The US Federal Reserve reported that, in 2018, PINs secured 31.4 billion transactions in the US, with an overall worth of US$ 1.19 trillion. One well-known attack type involves the use of cameras to spy on the ATM PIN pad during PIN entry. Countermeasures include covering the PIN pad with a shield or with the other hand while typing. Although this protects PINs from visual attacks, acoustic emanations from the PIN pad itself open the door for another attack type. In this paper, we show the feasibility of an acoustic side-channel attack (called PinDrop ) to reconstruct PINs by profiling acoustic signatures of individual keys of a PIN pad. We demonstrate the practicality of PinDrop via two sets of data collection experiments involving two commercially available metal PIN pad models and 58 participants who entered a total of 5,800 5-digit PINs. We simulated two realistic attack scenarios: (1) a microphone placed near the ATM (0.3 m away) and (2) a real-time attacker (with a microphone) standing in the queue at a common courtesy distance of 2 m. In the former case, we show that PinDrop recovers 96% of 4-digit, and up to 94% of 5-digits, PINs. Whereas, at 2 m away, it recovers up to 57% of 4-digit, and up to 39% of 5-digit PINs in three attempts. We believe that these results are both significant and worrisome.

The identification of emotions conveyed by faces as genuine or not is a topic under-explored. While some controversial insights are available for happiness (where genuine happiness is supposed to create crow's feet around the eyes), nothing is known regarding other emotions. This topic is important as human beings are known to perform around the chance level to identify genuine emotions. It is thus pivotal to identify the relevant features for the correct identification of a facial emotion as genuine. To this aim, we capitalized on explainable artificial intelligence (XAI), characterized by a superior ability to detect subtle patterns in the data. Two different XAI models were created and applied to a dataset including 50 participants displaying both genuine and not genuine emotions. Results revealed that ML models achieved high accuracies in genuineness discrimination (78 % of average) while being robust and interpretable. The robustness of the results, ensured by their stability across different models and experimental conditions, is critical to improving the generalizability of the results. Our XAI algorithms provided interpretable results as they correctly identified facial muscle movements that are critical for the classification of emotions as genuine or fake.

Automated Teller Machines (ATMs) represent the most used system for withdrawing cash. The European Central Bank reported more than 11 billion cash withdrawals and loading/unloading transactions on the European ATMs in 2019. Although ATMs have undergone various technological evolutions, Personal Identification Numbers (PINs) are still the most common authentication method for these devices. Unfortunately, the PIN mechanism is vulnerable to shoulder-surfing attacks performed via hidden cameras installed near the ATM to catch the PIN pad. To overcome this problem, people get used to covering the typing hand with the other hand. While such users probably believe this behavior is safe enough to protect against mentioned attacks, there is no clear assessment of this countermeasure in the scientific literature. This paper proposes a novel attack to reconstruct PINs entered by victims covering the typing hand with the other hand. We consider the setting where the attacker can access an ATM PIN pad of the same brand/model as the target one. Afterward, the attacker uses that model to infer the digits pressed by the victim while entering the PIN. Our attack owes its success to a carefully selected deep learning architecture that can infer the PIN from the typing hand position and movements. We run a detailed experimental analysis including 58 users. With our approach, we can guess 30% of the 5-digit PINs within three attempts - the ones usually allowed by ATM before blocking the card. We also conducted a survey with 78 users that managed to reach an accuracy of only 7.92% on average for the same setting. Finally, we evaluate a shielding countermeasure that proved to be rather inefficient unless the whole keypad is shielded.

Social bots are computer algorithms able to produce content and interact with other users on social media autonomously, trying to emulate and possibly influence humans' behavior. Indeed, bots are largely employed for malicious purposes, like spreading disinformation and conditioning electoral campaigns. Nowadays, bots' capability of emulating human behaviors has become increasingly sophisticated, making their detection harder. In this paper, we aim at recognizing bot-driven accounts by evaluating the consistency of users' writing style over time. In particular, we leverage the intuition that while bots compose posts according to fairly deterministic processes, humans are influenced by subjective factors (e.g., emotions) that can alter their writing style. To verify this assumption, by using stylistic consistency indicators, we characterize the writing style of more than 12,000 among bot-driven and human-operated Twitter accounts and find that statistically significant differences can be observed between the different types of users. Thus, we evaluate the effectiveness of different machine learning (ML) algorithms based on stylistic consistency features in discerning between human-operated and bot-driven Twitter accounts and show that the experimented ML algorithms can achieve high performance (i.e., F-measure values up to 98%) in social bot detection tasks.

Malingered responses to psychological testing are frequent when monetary incentives or other forms of rewards are at stake. Psychological symptoms are usually identified through clinical questionnaires which, however, may be easily inflated by malingered responses (fake-bad). A fake-bad response style is usually identified through specialized scales embedded in the personality questionnaires, but no procedure is currently available that reconstructs honest responses from malingered responses. In this paper, we present a technique for the Millon (MCMI-III) questionnaire a widely used test for investigating psychopathology. This technique detects malingered MCMI-III profiles (malingering detector) and removes the intentionally inflated test results (malingering remover). We demonstrate that by applying machine learning to the validity scales of MCMI-III we can discriminate between malingerer and honest profiles with 90% accuracy. Moreover, our results show that by applying regression models to malingerer tests, we are able to well reconstruct the original honest profile. Our models decrease the RMSE (Root Mean Square Error) of the reconstruction up to 19% compared to base correction procedures. Finally, applying the malingering detector to the reconstructed scales, we show that only 9% were classified as malingerers, demonstrating the validity of the proposed approach.

The growth of social media and the people interconnection led to the digitalization of communication. Nowadays the most influential politicians or scientific communicators use the media to disseminate news or decisions. However, such communications media can be used maliciously to spread the so-called fake-news in order to polarise public opinion or to deny scientific theories. It is therefore important to develop intelligent and accurate techniques in order to identify the spreading of fake-news. In this paper, we describes the methodology regarding our participation in the PAN@CLEF Profiling Fake News Spreaders on Twitter competition. We propose a supervised Machine-Learning (ML) based framework to profile fake-news spreaders. Our method relies on the combination of Big Five personality and stylometric features. Finally, we evaluate our framework detection capabilities and performance with different ML models on a tweeter dataset in both English and Spanish languages.

Keystroke dynamics has been recently proved to be an effective behavioral measure to detect subjects who provide false demographic information in online contexts. However, current techniques still suffer from some limits that restrict their practical application, such as the use of errors as a key feature to train the lie detectors and the absence of normalized features. Here, an extension of a keystroke dynamics technique, which was recently proposed to detect faked identities, is reported with the goal to overcome these limitations. Using a Quadratic Discriminant Analysis an accuracy up to 92% in the identification of faked identities has been reached, even if errors were excluded from predictors and normalized features were included. The classification model performs similarly to those previously proposed, with a slightly lower accuracy (−3%) but overcoming their important practical limitations.

Personal Identification Numbers (PINs) are widely used as the primary authentication method for Automated Teller Machines (ATMs) and Point of Sale (PoS). ATM and PoS typically mitigate attacks including shoulder-surfing by displaying dots on their screen rather than PIN digits, and by obstructing the view of the keypad. In this paper, we explore several sources of information leakage from common ATM and PoS installations that the adversary can leverage to reduce the number of attempts necessary to guess a PIN. Specifically, we evaluate how the adversary can leverage audio feedback generated by a standard ATM keypad to infer accurate inter-keystroke timing information, and how these timings can be used to improve attacks based on the observation of the user’s typing behavior, partial PIN information, and attacks based on thermal cameras. Our results show that inter-keystroke timings can be extracted from audio feedback far more accurately than from previously explored sources (e.g., videos). In our experiments, this increase in accuracy translated to a meaningful increase in guessing performance. Further, various combinations of these sources of information allowed us to guess between 44% and 89% of the PINs within 5 attempts. Finally, we observed that based on the type of information available to the adversary, and contrary to common knowledge, uniform PIN selection is not necessarily the best strategy. We consider these results relevant and important, as they highlight a real threat to any authentication system that relies on PINs.

Recently, the increasing spread of Online Social Networks (OSNs) provided an unprecedented opportunity of analysing online traces of human behaviour to get insight on individuals and society. Among the others, the possibility of predicting users' political orientation relying on data extracted from OSNs received growing attention. In this study, we introduce and make publicly available a dataset composed of 6.685 unique Twitter users and 9.593.055 Tweets. Differently from most of the dataset currently available in the literature, here, each user was manually labeled according to their political orientation by a pool of human judges, using strict inclusion criteria. Further, we address the feasibility of the automatic classification of Italian Twitter users' political orientation based on their Tweets content. Our analysis focuses first on implementing a series of classifiers with the aim of predicting users' political preference as right- or left-oriented. The built models were then evaluated for inferring the political orientation of those users supporting 'Movimento 5 Stelle' (M5S), an Italian political party with a still unclear political leaning. Results show high performances on the left-right classification task, with accuracy rates up to 93%. Finally, classification performances obtained on M5S supporters and possible applications of our findings are discussed.