Artificial intelligence (AI) synthesized faces—so called deepfake images—have been increasingly used for malicious intent and have resulted in prominently adverse impact. Because online users must contend with discerning fake from real, great emphasis has been placed on enhancing human detection of deepfake images. We conducted an online human-subject study (N=237), investigating the effect of three training strategies (explicit training with visible artifacts in synthetic faces, implicit training with experiencing the generation of synthetic faces using real human faces, and a combination of both artifact and generation) on participants’ detection of synthetic faces generated by the state-of-the-art StyleGAN techniques. Comparing participants’ deepfake detection across three phases (baseline in phase 1 without any training, phase 2 after one training session, and phase 3 after the other training session), we found that all training strategies effectively enhanced participants’ detection of AI-synthesized faces and their decision confidence. We also explored factors that impact participants’ learning and decision-making of deepfake detection. Responses to the open-ended question revealed that participants developed generalized strategies and utilized artifacts beyond the training. Our quantitative and qualitative results provide nuanced insights into the promises and limitations of the training strategies. In addition to advancing theoretical understanding of human training in the context of deepfake image detection, our study findings hold practical implications for interface design.

Recently, deepfake techniques have been adopted by real-world adversaries to fabricate believable personas (posing as experts or insiders) in disinformation campaigns to promote false narratives and deceive the public. In this paper, we investigate how fake personas influence the user perception of the disinformation shared by such accounts. Using Twitter as an exemplary platform, we conduct a user study (N=417) where participants read tweets of fake news with (and without) the presence of the tweet authors' profiles. Our study examines and compares three types of fake profiles: deepfake profiles, profiles of relevant organizations, and simple bot profiles. Our results highlight the significant impact of deepfake and organization profiles on increasing the perceived information accuracy of and engagement with fake news. Moreover, deepfake profiles are rated as significantly more real than other profile types. Finally, we observe that users may like/reply/share a tweet even though they believe it was inaccurate (e.g., for fun or truth-seeking), which could further disseminate false information. We then discuss the implications of our findings and directions for future research.

Machine learning based phishing website detectors (ML-PWD) are a critical part of today's anti-phishing solutions in operation. Unfortunately, ML-PWD are prone to adversarial evasions, evidenced by both academic studies and analyses of real-world adversarial phishing webpages. However, existing works mostly focused on assessing adversarial phishing webpages against ML-PWD, while neglecting a crucial aspect: investigating whether they can deceive the actual target of phishing---the end users. In this paper, we fill this gap by conducting two user studies (n=470) to examine how human users perceive adversarial phishing webpages, spanning both synthetically crafted ones (which we create by evading a state-of-the-art ML-PWD) as well as real adversarial webpages (taken from the wild Web) that bypassed a production-grade ML-PWD. Our findings confirm that adversarial phishing is a threat to both users and ML-PWD, since most adversarial phishing webpages have comparable effectiveness on users w.r.t. unperturbed ones. However, not all adversarial perturbations are equally effective. For example, those with added typos are significantly more noticeable to users, who tend to overlook perturbations of higher visual magnitude (such as replacing the background). We also show that users' self-reported frequency of visiting a brand's website has a statistically negative correlation with their phishing detection accuracy, which is likely caused by overconfidence. We release our resources.

Federal funding agencies and industry entities are seeking innovative approaches to address the ever-growing cybersecurity crisis. Increasingly, numerous cybersecurity thought leaders are indicating that Artificial Intelligence (AI)-enabled analytics can help tackle key cybersecurity tasks and deploy defenses. This half-day workshop, co-located with ACM KDD, sought to attain significant research contributions to various aspects of AI-enabled analytics for cybersecurity applications and deployable defense solutions from academics and practitioners. This workshop was a joint workshop of the 2021 AI-enabled Cybersecurity Analytics and 2021 International Workshop on Deployable Machine Learning for Security Defense. As such, we developed an interdisciplinary Program Committee with significant experience in various aspects of AI, cybersecurity, and/or deployable defense.

Email authentication protocols such as SPF, DKIM, and DMARC are used to detect spoofing attacks, but they face key challenges when handling email forwarding scenarios. Recently in 2019, a new Authenticated Received Chain (ARC) protocol was introduced to support mail forwarding applications to preserve the authentication records. After 2 years, it is still not well understood how ARC is implemented, deployed, and configured in practice. In this paper, we perform an empirical analysis on ARC usage and examine how it affects spoofing detection decisions on popular email provides that support ARC. After analyzing an email dataset of 600K messages, we show that ARC is not yet widely adopted, but it starts to attract adoption from major email providers (e.g., Gmail, Outlook). Our controlled experiment shows that most email providers’ ARC implementations are done correctly. However, some email providers (Zoho) have misinterpreted the meaning of ARC results, which can be exploited by spoofing attacks. Finally, we empirically investigate forwarding-based “Hide My Email” services offered by iOS 15 and Firefox, and show their implementations break ARC and can be leveraged by attackers to launch more successful spoofing attacks against otherwise well-configured email receivers (e.g., Gmail).

Different techniques have been recommended to detect fraudulent responses in online surveys, but little research has been taken to systematically test the extent to which they actually work in practice. In this paper, we conduct an empirical evaluation of 22 anti-fraud tests in two complementary online surveys. The first survey recruits Rust programmers on public online forums and social media networks. We find that fraudulent respondents involve both bot and human characteristics. Among different anti-fraud tests, those designed based on domain knowledge are the most effective. By combining individual tests, we can achieve a detection performance as good as commercial techniques while making the results more explainable. To explore these tests under a broader context, we ran a different survey on Amazon Mechanical Turk (MTurk). The results show that for a generic survey without requiring users to have any domain knowledge, it is more difficult to distinguish fraudulent responses. However, a subset of tests still remain effective.

The MLHat workshop aims to bring together academic researchers and industry practitioners to discuss the open challenges, potential solutions, and best practices to deploy machine learning at scale for security defense. The workshop will discuss related topics from both defender perspectives (white-hat) and the attacker perspectives (black-hat). We call the workshop MLHats, to serve as a place for people who are interested in using machine learning to solve practical security problems. The workshop will focus on defining new machine learning paradigms under various security application contexts and identifying exciting new future research directions. At the same time, the workshop will also have a strong industry presence to provide insights into the challenges in deploying and maintaining machine learning models and the much-needed discussion on the capabilities that the state-of-the-arts failed to provide.

Understanding the linkability of online user identifiers (IDs) is critical to both service providers (for business intelligence) and individual users (for assessing privacy risks). Existing methods are designed to match IDs across two services but face key challenges of matching multiple services in practice, particularly when users have multiple IDs per service. In this article, we propose a novel system to link IDs across multiple services by exploring the spatial-temporal features of user activities, of which the core idea is that the same user's online IDs are more likely to repeatedly appear at the same location. Specifically, we first utilize a contact graph to capture the “co-location” of all IDs across multiple services. Based on this graph, we propose a set-wise matching algorithm to discover candidate ID sets and use Bayesian inference to generate confidence scores for candidate ranking, which is proved to be optimal. We evaluate our system using two real-world ground-truth datasets from an Internet service provider (4 services, 815K IDs) and Twitter-Foursquare (2 services, 770 IDs). Extensive results show that our system significantly outperforms the state-of-the-art algorithms in accuracy (AUC is higher by 0.1–0.2), and it is highly robust against data quality, matching order, and number of services.

Voice Personal Assistant (VPA) systems such as Amazon Alexa and Google Home have been used by tens of millions of households. Recent work demonstrated proof-of-concept attacks against their voice interface to invoke unintended applications or operations. However, there is still a lack of empirical understanding of what type of third-party applications that VPA systems support, and what consequences these attacks may cause. In this paper, we perform an empirical analysis of the third-party applications of Amazon Alexa and Google Home to systematically assess the attack surfaces. A key methodology is to characterize a given application by classifying the sensitive voice commands it accepts. We develop a natural language processing tool that classifies a given voice command from two dimensions: (1) whether the voice command is designed to insert action or retrieve information; (2) whether the command is sensitive or nonsensitive. The tool combines a deep neural network and a keyword-based model, and uses Active Learning to reduce the manual labeling effort. The sensitivity classification is based on a user study (N=404) where we measure the perceived sensitivity of voice commands. A ground-truth evaluation shows that our tool achieves over 95% of accuracy for both types of classifications. We apply this tool to analyze 77,957 Amazon Alexa applications and 4,813 Google Home applications (198,199 voice commands from Amazon Alexa, 13,644 voice commands from Google Home) over two years (2018-2019). In total, we identify 19,263 sensitive “action injection” commands and 5,352 sensitive “information retrieval” commands. These commands are from 4,596 applications (5.55% out of all applications), most of which belong to the “smart home” category. While the percentage of sensitive applications is small, we show the percentage is increasing over time from 2018 to 2019.

Understanding the linkability of online user identifiers (IDs) is critical to both service providers (for business intelligence) and individual users (for assessing privacy risks). Existing methods are designed to match IDs across two services, but face key challenges of matching multiple services in practice, particularly when users have multiple IDs per service. In this paper, we propose a novel system to link IDs across multiple services by exploring the spatial-temporal locality of user activities. The core idea is that the same user's online IDs are more likely to repeatedly appear at the same location. Specifically, we first utilize a contact graph to capture the “co-location” of all IDs across multiple services. Based on this graph, we propose a set-wise matching algorithm to discover candidate ID sets, and use Bayesian inference to generate confidence scores for candidate ranking, which is proved to be optimal. We evaluate our system using two real-world ground-truth datasets from an ISP (4 services, 815K IDs) and Twitter-Foursquare (2 services, 770 IDs). Extensive results show that our system significantly outperforms the state-of-the-art algorithms in accuracy (AUC is higher by 0.1–0.2), and it is highly robust against matching order and number of services.

Social media is often viewed as a sensor into various societal events such as disease outbreaks, protests, and elections. We describe the use of social media as a crowdsourced sensor to gain insight into ongoing cyber-attacks. Our approach detects a broad range of cyber-attacks (e.g., distributed denial of service (DDoS) attacks, data breaches, and account hijacking) in a weakly supervised manner using just a small set of seed event triggers and requires no training or labeled samples. A new query expansion strategy based on convolution kernels and dependency parses helps model semantic structure and aids in identifying key event characteristics. Through a large-scale analysis over Twitter, we demonstrate that our approach consistently identifies and encodes events, outperforming existing methods.

Today's video streaming market is crowded with various content providers (CPs). For individual CPs, understanding user behavior, in particular how users migrate among different CPs, is crucial for improving users' on-site experience and the CP's chance of success. In this paper, we take a data-driven approach to analyze and model user migration behavior in video streaming, i.e., users switching content provider during active sessions. Based on a large ISP dataset over two months (6 major content providers, 3.8 million users, and 315 million video requests), we study common migration patterns and reasons of migration. We find that migratory behavior is prevalent: 66% of users switch CPs with an average switching frequency of 13%. In addition, migration behaviors are highly diverse: regardless large or small CPs, they all have dedicated groups of users who like to switch to them for certain types of videos. Regarding reasons of migration, we find CP service quality rarely causes migration, while a few popular videos play a bigger role. Nearly 60% of cross-site migrations are landed to 0.14% top videos. Finally, we validate our findings by building an accurate regression model to predict user migration frequency, and discuss the implications of our results to CPs.

Digital location traces can help build insights about how citizens experience their cities, but also offer personalized products and experiences to them. Even as data abound, though, building an accurate picture about citizen whereabouts is not always straightforward, due to noisy or incomplete data.

It is often difficult to separate the highly capable “experts” from the average worker in crowdsourced systems. This is especially true for challenge application domains that require extensive domain knowledge. The problem of stock analysis is one such domain, where even the highly paid, well-educated domain experts are prone to make mistakes. As an extremely challenging problem space, the “wisdom of the crowds” property that many crowdsourced applications rely on may not hold. In this article, we study the problem of evaluating and identifying experts in the context of SeekingAlpha and StockTwits, two crowdsourced investment services that have recently begun to encroach on a space dominated for decades by large investment banks. We seek to understand the quality and impact of content on collaborative investment platforms, by empirically analyzing complete datasets of SeekingAlpha articles (9 years) and StockTwits messages (4 years). We develop sentiment analysis tools and correlate contributed content to the historical performance of relevant stocks. While SeekingAlpha articles and StockTwits messages provide minimal correlation to stock performance in aggregate, a subset of experts contribute more valuable (predictive) content. We show that these authors can be easily identified by user interactions, and investments based on their analysis significantly outperform broader markets. This effectively shows that even in challenging application domains, there is a secondary or indirect wisdom of the crowds. Finally, we conduct a user survey that sheds light on users’ views of SeekingAlpha content and stock manipulation. We also devote efforts to identify potential manipulation of stocks by detecting authors controlling multiple identities.

The next generation of Internet services is driven by users and user-generated content. The complex nature of user behavior makes it highly challenging to manage and secure online services. On one hand, service providers cannot effectively prevent attackers from creating large numbers of fake identities to disseminate unwanted content (e.g., spam). On the other hand, abusive behavior from real users also poses significant threats (e.g., cyberbullying). In this article, we propose clickstream models to characterize user behavior in large online services. By analyzing clickstream traces (i.e., sequences of click events from users), we seek to achieve two goals: (1) detection: to capture distinct user groups for the detection of malicious accounts, and (2) understanding: to extract semantic information from user groups to understand the captured behavior. To achieve these goals, we build two related systems. The first one is a semisupervised system to detect malicious user accounts (Sybils). The core idea is to build a clickstream similarity graph where each node is a user and an edge captures the similarity of two users’ clickstreams. Based on this graph, we propose a coloring scheme to identify groups of malicious accounts without relying on a large labeled dataset. We validate the system using ground-truth clickstream traces of 16,000 real and Sybil users from Renren, a large Chinese social network. The second system is an unsupervised system that aims to capture and understand the fine-grained user behavior. Instead of binary classification (malicious or benign), this model identifies the natural groups of user behavior and automatically extracts features to interpret their semantic meanings. Applying this system to Renren and another online social network, Whisper (100K users), we help service providers identify unexpected user behaviors and even predict users’ future actions. Both systems received positive feedback from our industrial collaborators including Renren, LinkedIn, and Whisper after testing on their internal clickstream data.

Efforts such as Wikipedia have shown the ability of user communities to collect, organize and curate information on the Internet. Recently, a number of question and answer (Q&A) sites have successfully built large growing knowledge repositories, each driven by a wide range of questions and answers from its users community. While sites like Yahoo Answers have stalled and begun to shrink, one site still going strong is Quora, a rapidly growing service that augments a regular Q&A system with social links between users. Despite its success, however, little is known about what drives Quora's growth, and how it continues to connect visitors and experts to the right questions as it grows.

Popular Internet services in recent years have shown that remarkable things can be achieved by harnessing the power of the masses using crowd-sourcing systems. However, crowd-sourcing systems can also pose a real challenge to existing security mechanisms deployed to protect Internet services. Many of these security techniques rely on the assumption that malicious activity is generated automatically by automated programs. Thus they would perform poorly or be easily bypassed when attacks are generated by real users working in a crowd-sourcing system. Through measurements, we have found surprising evidence showing that not only do malicious crowd-sourcing systems exist, but they are rapidly growing in both user base and total revenue. We describe in this paper a significant effort to study and understand these "crowdturfing" systems in today's Internet. We use detailed crawls to extract data about the size and operational structure of these crowdturfing systems. We analyze details of campaigns offered and performed in these sites, and evaluate their end-to-end effectiveness by running active, benign campaigns of our own. Finally, we study and compare the source of workers on crowdturfing sites in different countries. Our results suggest that campaigns on these systems are highly effective at reaching users, and their continuing growth poses a concrete threat to online communities both in the US and elsewhere.

Abstract The rapid growth of the non‐English‐speaking Internet population has created a need for better searching and browsing capabilities in languages other than English. However, existing search engines may not serve the needs of many non‐English‐speaking Internet users. In this paper, we propose a generic and integrated approach to searching and browsing the Internet in a multilingual world. Based on this approach, we have developed the Chinese Business Intelligence Portal (CBizPort), a meta‐search engine that searches for business information of mainland China, Taiwan, and Hong Kong. Additional functions provided by CBizPort include encoding conversion (between Simplified Chinese and Traditional Chinese), summarization, and categorization. Experimental results of our user evaluation study show that the searching and browsing performance of CBizPort was comparable to that of regional Chinese search engines, and CBizPort could significantly augment these search engines. Subjects' verbal comments indicate that CBizPort performed best in terms of analysis functions, cross‐regional searching, and user‐friendliness, whereas regional search engines were more efficient and more popular. Subjects especially liked CBizPort's summarizer and categorizer, which helped in understanding search results. These encouraging results suggest a promising future of our approach to Internet searching and browsing in a multilingual world.