The European General Data Protection Regulation (GDPR) mandates a data controller (e.g., an app developer) to provide all information specified in Articles (Arts.) 13 and 14 to data subjects (e.g., app users) regarding how their data are being processed and what are their rights. While some studies have started to detect the fulfillment of GDPR requirements in a privacy policy, their exploration only focused on a subset of mandatory GDPR requirements. In this paper, our goal is to explore the state of GDPR-completeness violations in mobile apps' privacy policies. To achieve our goal, we design the PolicyChecker framework by taking a rule and semantic role based approach. PolicyChecker automatically detects completeness violations in privacy policies based not only on all mandatory GDPR requirements but also on all if-applicable GDPR requirements that will become mandatory under specific conditions. Using PolicyChecker, we conduct the first large-scale GDPR-completeness violation study on 205,973 privacy policies of Android apps in the UK Google Play store. PolicyChecker identified 163,068 (79.2%) privacy policies containing data collection statements; therefore, such policies are regulated by GDPR requirements. However, the majority (99.3%) of them failed to achieve the GDPR-completeness with at least one unsatisfied requirement; 98.1% of them had at least one unsatisfied mandatory requirement, while 73.0% of them had at least one unsatisfied if-applicable requirement logic chain. We conjecture that controllers' lack of understanding of some GDPR requirements and their poor practices in composing a privacy policy can be the potential major causes behind the GDPR-completeness violations. We further discuss recommendations for app developers to improve the completeness of their apps' privacy policies to provide a more transparent personal data processing environment to users.

Mobile and web apps are increasingly relying on the data generated or provided by users such as from their uploaded documents and images. Unfortunately, those apps may raise significant user privacy concerns. Specifically, to train or adapt their models for accurately processing huge amounts of data continuously collected from millions of app users, app or service providers have widely adopted the approach of crowdsourcing for recruiting crowd workers to manually annotate or transcribe the sampled ever-changing user data. However, when users' data are uploaded through apps and then become widely accessible to hundreds of thousands of anonymous crowd workers, many human-in-the-loop related privacy questions arise concerning both the app user community and the crowd worker community. In this paper, we propose to investigate the privacy risks brought by this significant trend of large-scale crowd-powered processing of app users' data generated in their daily activities. We consider the representative case of receipt scanning apps that have millions of users, and focus on the corresponding receipt transcription tasks that appear popularly on crowdsourcing platforms. We design and conduct an app user survey study (n=108) to explore how app users perceive privacy in the context of using receipt scanning apps. We also design and conduct a crowd worker survey study (n=102) to explore crowd workers' experiences on receipt and other types of transcription tasks as well as their attitudes towards such tasks. Overall, we found that most app users and crowd workers expressed strong concerns about the potential privacy risks to receipt owners, and they also had a very high level of agreement with the need for protecting receipt owners' privacy. Our work provides insights on app users' potential privacy risks in crowdsourcing, and highlights the need and challenges for protecting third party users' privacy on crowdsourcing platforms. We have responsibly disclosed our findings to the related crowdsourcing platform and app providers.

Natural Language Processing (NLP) models are often vulnerable to semantics-preserving adversarial attacks. That is, they make different semantic predictions on input instances with similar content and semantics. However, it remains unclear to which extent modern NLP models are vulnerable to content-preserving and semantics-flipping (CPSF) adversarial attacks. That is, they would make the same semantic prediction on input instances with similar content but flipped semantics. Attackers can use either semantics-preserving or CPSF adversarial examples to create misunderstanding between humans and models, and incur severe consequences in real-world applications. However, this equally important problem on CPSF adversarial examples has not been studied by researchers yet. In this paper, we perform the first study to investigate CPSF adversarial examples and propose CPSF adversarial attacks to reveal this new type of vulnerability of NLP models. We develop a two-stage approach to generate CPSF adversarial examples. Our experiments on two types of NLP tasks, sentiment analysis and textual entailment, demonstrate that CPSF adversarial examples can successfully fool victim models while preserving the same content with flipped semantics to humans. We further validate the good transferability of CPSF adversarial examples on NLP services of Microsoft and Google. Moreover, we demonstrate that adversarial training can to a meaningful extent mitigate CPSF adversarial attacks. Overall, our work implies that researchers need to improve NLP models' robustness against CPSF adversarial attacks that uniquely exploit the blind spots where NLP models are too insensitive to even big changes in semantics.

Web tracking and advertising (WTA) nowadays are ubiquitously performed on the web, continuously compromising users’ privacy. Existing defense solutions, such as widely deployed blocking tools based on filter lists and alternative machine learning based solutions proposed in prior research, have limitations in terms of accuracy and effectiveness. In this work, we propose WTAGRAPH, a web tracking and advertising detection framework based on Graph Neural Networks (GNNs). We first construct an attributed homogenous multi-graph (AHMG) that represents HTTP network traffic, and formulate web tracking and advertising detection as a task of GNN-based edge representation learning and classification in AHMG. We then design four components in WTAGRAPH so that it can (1) collect HTTP network traffic, DOM, and JavaScript data, (2) construct AHMG and extract corresponding edge and node features, (3) build a GNN model for edge representation learning and WTA detection in the transductive learning setting, and (4) use a pre-trained GNN model for WTA detection in the inductive learning setting. We evaluate WTAGRAPH on a dataset collected from Alexa Top 10K websites, and show that WTAGRAPH can effectively detect WTA requests in both transductive and inductive learning settings. Manual verification results indicate that WTAGRAPH can detect new WTA requests that are missed by filter lists and recognize non-WTA requests that are mistakenly labeled by filter lists. Our ablation analysis, evasion evaluation, and real-time evaluation show that WTAGRAPH can have a competitive performance with flexible deployment options in practice.

Crowdsourcing is popular for large-scale data collection and labeling, but a major challenge is on detecting low-quality submissions. Recent studies have demonstrated that behavioral features of workers are highly correlated with data quality and can be useful in quality control. However, these studies primarily leveraged coarsely extracted behavioral features, and did not further explore quality control at the fine-grained level, i.e., the annotation unit level. In this paper, we investigate the feasibility and benefits of using fine-grained behavioral features, which are the behavioral features finely extracted from a worker's individual interactions with each single unit in a subtask, for quality control in crowdsourcing. We design and implement a framework named Fine-grained Behavior-based Quality Control (FBQC) that specifically extracts fine-grained behavioral features to provide three quality control mechanisms: (1) quality prediction for objective tasks, (2) suspicious behavior detection for subjective tasks, and (3) unsupervised worker categorization. Using the FBQC framework, we conduct two real-world crowdsourcing experiments and demonstrate that using fine-grained behavioral features is feasible and beneficial in all three quality control mechanisms. Our work provides clues and implications for helping job requesters or crowdsourcing platforms to further achieve better quality control.

Attention check questions have become commonly used in online surveys published on popular crowdsourcing platforms as a key mechanism to filter out inattentive respondents and improve data quality. However, little research considers the vulnerabilities of this important quality control mechanism that can allow attackers including irresponsible and malicious respondents to automatically answer attention check questions for efficiently achieving their goals. In this paper, we perform the first study to investigate such vulnerabilities, and demonstrate that attackers can leverage deep learning techniques to pass attention check questions automatically. We propose AC-EasyPass, an attack framework with a concrete model, that combines convolutional neural network and weighted feature reconstruction to easily pass attention check questions. We construct the first attention check question dataset that consists of both original and augmented questions, and demonstrate the effectiveness of AC-EasyPass. We explore two simple defense methods, adding adversarial sentences and adding typos, for survey designers to mitigate the risks posed by AC-EasyPass; however, these methods are fragile due to their limitations from both technical and usability perspectives, underlining the challenging nature of defense. We hope our work will raise sufficient attention of the research community towards developing more robust attention check mechanisms. More broadly, our work intends to prompt the research community to seriously consider the emerging risks posed by the malicious use of machine learning techniques to the quality, validity, and trustworthiness of crowdsourcing and social computing.

Non-negative sparse recovery refers to recovering non-negative sparse source signals from linear observations. This model arises naturally in many image processing applications such as super-resolution and image inpainting. In this paper, we propose two efficient neural networks for fast approximation of non-negative sparse recovery. We also derive upper bounds on network sizes measured by the numbers of layers and neurons to achieve a specified approximation error. Numerical experiments demonstrate the effectiveness and robustness of the proposed networks and show their potential in solving more complicated signal recovery problems with non-stationary transformation process and noisy observation.