Abstract With the rise of Web 2.0 platforms such as online social media, people’s private information, such as their location, occupation and even family information, is often inadvertently disclosed through online discussions. Therefore, it is important to detect such unwanted privacy disclosures to help alert people affected and the online platform. In this paper, privacy disclosure detection is modeled as a multi-label text classification (MLTC) problem, and a new privacy disclosure detection model is proposed to construct an MLTC classifier for detecting online privacy disclosures. This classifier takes an online post as the input and outputs multiple labels, each reflecting a possible privacy disclosure. The proposed presentation method combines three different sources of information, the input text itself, the label-to-text correlation and the label-to-label correlation. A double-attention mechanism is used to combine the first two sources of information, and a graph convolutional network is employed to extract the third source of information that is then used to help fuse features extracted from the first two sources of information. Our extensive experimental results, obtained on a public dataset of privacy-disclosing posts on Twitter, demonstrated that our proposed privacy disclosure detection method significantly and consistently outperformed other state-of-the-art methods in terms of all key performance indicators.

Mis-and disinformation online have become a major societal problem as major sources of online harms of different kinds.One common form of mis-and disinformation is outof-context (OOC) information, where different pieces of information are falsely associated, e.g., a real image combined with a false textual caption or a misleading textual description.Although some past studies have attempted to defend against OOC mis-and disinformation through external evidence, they tend to disregard the role of different pieces of evidence with different stances.Motivated by the intuition that the stance of evidence represents a bias towards different detection results, we propose a stance extraction network (SEN) that can extract the stances of different pieces of multi-modal evidence in a unified framework.Moreover, we introduce a support-refutation score calculated based on the co-occurrence relations of named entities into the textual SEN.Extensive experiments on a public large-scale dataset demonstrated that our proposed method outperformed the state-ofthe-art baselines, with the best model achieving a performance gain of 3.2% in accuracy.* Corresponding co-authors 1 In the literature the terms "misinformation" and "disinformation" often have inconsistent definitions.In our work, we adopt the more established definitions by the United Nations (https://www.undp.org/eurasia/dis/ misinformation): misinformation refers to information that is false but not created with the intention of causing harm and disinformation to information that is false and deliberately created to cause harm.Our work can be applied to both mis-and disinformation, so we will mostly use the term "mis-/disinformation".

Multi-label text classification, which aims to predict the relevant labels for each given document, is one of the fundamental tasks of natural language processing. Recent studies have utilized Transformer, which embeds texts and class labels into a joint space to capture the label correlation. However, existing methods tend to take up extra input length and ignore the significance of taxonomic hierarchy. For this reason, we introduce a label correlation enhanced decoder (LED) for multi-label text classification. LED predicts the presence or absence of class labels in parallel with label representation and captures label correlation through multi-task learning. In addition, we propose a hierarchy-aware mask to capture the hierarchical dependency between labels. Comprehensive experiments on four benchmark datasets show that LED outperforms the state-of-the-art baselines. Detailed analysis validates the effectiveness of our proposed method.

Anonymous users on Twitter are highly related to network hazards such as cyberbullying, fraud, rumor spread, etc. Given the inefficient and expensive manual review approach, an automated method to predict account anonymity is urgently needed. However, there are three challenges when employing deep learning algorithms to detect account anonymity: the lack of open-source public datasets, tedious manual collection and labeling for training data, and abundant types of information that the existing single-modal methods are not competent for. Our approach includes an automated labeling system for data collection and a multi-modal method for account anonymity detection. The labeling system is based on name extraction and identity verification and, to ensure the quality of the dataset, web mining and face similarity measurement are applied. We establish a dataset containing 20133 accounts and make it available to the public. The multi-modal method exploits the visual, textual, and numerical features extracted with ViT, BERT, and MLP, and uses the Transformer and MLP for feature fusion. Experimental results prove the effectiveness of the feature fusion and the accuracy of classification with accuracy of 86.21% and 92.46%.

Tabular data, with an exceedingly effective data structure, can give us a more intuitive visual display. To under-stand well and make use of the spatial and logic dependencies of it, we propose an end-to-end, graph-based table reconstructed network, namely GTRNet, in this paper. Our model works differently from most existing models which treat tables as either a markup sequence problem or a graph structure of rows and columns. It can utilize a table as input and extract its features in the text, image and position coordinate to predict the dependencies of the text instances and well distinguish the spatial relationship to infer whether multiple text segments belong to the same merged cell. Optimized along with this network, we can then restore the structure of this table. Moreover, we also create a new Chinese benchmark dataset GraphTable for this task to tackle complex challenges on the table. The competitive results on ICDAR-2013, GraphTable, SciTSR and FinTab benchmarks further confirm the great effectiveness of GTRN et.

While many online platforms bring great benefits to their users by allowing user-generated content, they have also facilitated generation and spreading of harmful content such as rumors. Researcher have proposed different rumor detection methods based on features extracted from the original post and/or associated comments, but how comments affect the performance of such methods remains largely less understood. In this paper, we first propose a new BERT-based rumor detection method that can outperform other state-of-the-art methods, and then used it to study the role of comments in rumor detection. Our proposed method concatenates the original post and associated comments to form a single long text, which is then segmented into shorter chunks more suitable for BERT-based vectorization. Features extracted from all trunks are fed into a classifier based on an LSTM network or a transformer layer for the classification task. The experimental results on the PHEME and Ma-Weibo datasets proved the superior performance of our method. We conducted additional experiments on different settings of our proposed method to study different aspects of the role comments play in the rumor detection task. These additional experiments led to some very interesting findings, including the surprising result that fixed-length segmentation is better than natural segmentation, and the observation that including more comments can help improve the rumor detector’s performance. Some of these findings have profound operational implications for online platforms, e.g., commentators can contribute to rumor detection positively so online platforms can leverage the crowd intelligence to detect online rumors more effectively without applying overstrict content consensus policies.

The spread of terroristic images on the World Wide Web will bring significant risks to social security. Terroristic image detection technology can help image filtering. Otherwise, due to the lack of high-quality terrorist image dataset, deep learning based recognition methods have not been popularized. Furthermore, characteristics of occlusions and diversity of scenes, automatic approaches to terroristic content detection need to be well-designed. In this paper, a multi-model system is intended to detect for various types of terroristic content. For an input image, the system will locate and identify the terrorist organization leader or flag, determine whether the text information in the image belongs to terrorist slogan, and distinguish a terroristic picture from an ordinary one. For a terroristic image, the system will further detect sensitive objects, e.g. guns or flames, and subdivide the scene of brutal terrorist events. We build a terroristic dataset containing over 41,000 images for experimentation. The experiment result shows the accuracy of the multi-model terroristic content detection system.

It is of great practical significance to accurately model and analyze abnormal events in time series. For example, the identification of anomaly patterns on infrastructure sensor curves helps locate equipment failures. In this paper, we propose an unsupervised anomaly detection approach for time series, which can comprehensively consider both point anomalies and subsequence anomalies. We innovatively introduce RNN into the architecture of Adversarial Autoencoder to better analyze anomaly events based on overall relationship of time series. In addition, we innovatively apply the Outlier Exposure technique for the performance optimization of anomaly detector. Meanwhile, a WGAN-based method is utilized to generate anomaly datasets through normal distribution learning. Finally, we apply the proposed method for fraud detection on a financial statement dataset and intrusion detection on a network traffic dataset. Experimental results demonstrates that our model can comprehensively consider different anomaly types in time series, and achieve promising detection performance overall. In the experiment of fraud detection, the LSTM integrated AAE model achieves an F1 score of 0.810, while the Outlier Exposure enhanced model achieves an F1 score of 0.894. This indicates that our method can improve the performance of current audit systems and facilitate discovering malicious behaviors.

Within the research of scene text detection, some previous work has already achieved significant accuracy and efficiency. However, most of the work was generally done without considering about the implicit relationship between detection and eye movements. In this paper, we propose a new method for scene text detection especially for its refinement based on reinforcement learning. The idea of this method is inspired by Saccadic Eye Movements and Peripheral Vision. A saccade makes it possible for humans to orient the gaze to the location where a visual object has appeared. Peripheral vision gathers visual information of surroundings which provides supplement to foveal vision during gazing. We propose a simple pipeline, imitating the way human eyes do a saccade and collect peripheral information, to locate scene text roughly and to refine multi-scale vision field iteratively using reinforcement learning. For both training and evaluation, we use ICDAR2015 Challenge 4 dataset as a base and design several criteria to measure the feasibility of our work.

Information extraction from document collections is a fundamental and important step to understand, structure and analyze data. Many approaches with rules and deep learning based techniques have been applied on plain text, however, when it comes to document images, such demand still exists but becomes quite challenging without linguistic knowledge. In this paper, we propose an approach to extract required named entities (NEs) from document images by integrating the coordinate information from the detection and recognition stage into the contextual information of the BiLSTM-CRF model with an attention mechanism. We test this method on two real-world datasets. One is a Contract Dataset of Listed Companies, and the other is an Insurance Policy Dataset of our own. The result shows a combination of coordinates and context with attention leverages extraction in document images, opening up potential applications on such tasks.

In this paper, we propose an end-to-end trainable framework for scene text spotting which can handle text with arbitrary shapes. The proposed framework is called Word Segmentation Guided Characters Aggregation Net (WAC-Net), which consists of a shared convolutional backbone and two task-specific subnetworks. One subnetwork does word-level instance-aware segmentation (WSN) and the other does char-level detection and recognition (CDRN). The entire framework segments each word instance while detects and recognizes each character in one single forward pass. These two subnetworks are jointly trained by multi-task learning. At the inference stage, characters are aggregated into words guided by word instance segmentation results. Experiments are conducted on two datasets with arbitrary shapes, and the results demonstrate the effectiveness of the proposed method.

In this paper, we propose Double Supervised Network with Attention Mechanism (DSAN), a novel end-to-end trainable framework for scene text recognition. It incorporates one text attention module during feature extraction which enforces the model to focus on text regions and the whole framework is supervised by two branches. One supervision comes from context-level modelling branch and another comes from one extra supervision enhancement branch which aims at tackling inexplicit semantic information at character level. These two supervisions can benefit each other and yield better performance. The proposed approach can recognize text in arbitrary length and does not need any predefined lexicon. Our method achieves the current state-of-the-art results on three text recognition benchmarks: IIIT5K, ICDAR2013 and SVT reaching accuracy 88.6%, 92.3% and 84.1% respectively which suggests the effectiveness of the proposed method.

In this paper, we introduce a novel end-end framework for multi-oriented scene text detection from an instance-aware semantic segmentation perspective. We present Fused Text Segmentation Networks, which combine multi-level features during the feature extracting as text instance may rely on finer feature expression compared to general objects. It detects and segments the text instance jointly and simultaneously, leveraging merits from both semantic segmentation task and region proposal based object detection task. Not involving any extra pipelines, our approach surpasses the current state of the art on multi-oriented scene text detection benchmarks: ICDAR2015 Incidental Scene Text and MSRA-TD500 reaching Hmean 84.1 % and 82.0 % respectively. Morever, we report a baseline on total-text containing curved text which suggests effectiveness of the proposed approach.

Traditional file recovery methods rely on file system information, which are ineffective when file system information isn't available. File carving is a file recovery method that recovers files according to their structure and content without file system information, which is widely used in digital forensics. As the important carriers of digital information, multimedia files are important digital evidence. In this paper, a new multimedia file carving approach is proposed to improve the recovery accuracy of high entropy file fragments. The fragmented files can be recovered by a hierarchical carving process, including file header identification via entropy, file fragment type classification, and file reassembly via parallel unique path approach. A new file type classification method is constructed based on support vector machine, by using the features of BFD (byte frequency distribution) and ROC (rate of change). Four different datasets, such as DFRWS 2006/2007 challenge datasets, dataset simulating actual disk, dataset with randomly disordered fragments, and dataset with biomedical images, are employed in our experiments. The results show that JPEG recovery accuracy is improved greatly compared with that of Photo Rec tool. Our method performs best in the situation where the order of fragments is completely confusing.