Yueheng Mao, Min Yu, Gengwang Li, Jianguo Jiang, Gang Li, Meng Zhang, Zhen Xu, Weiqing Huang, Ming Liu. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Deepfake detectors have demonstrated vulnerability when faced with adversarial attacks. In real-world scenarios, attackers can generate adversarial examples that mislead detectors through query-based black-box attack methods. Recently, several defense methods against query-based adversarial attacks have achieved promising performance. However, their effectiveness significantly declines when applied to deepfake detection tasks. In this work, we propose a novel defense method specifically designed to counter adversarial attacks in deepfake detection scenarios. Unlike existing approaches, our method focuses on the high-frequency components of images to uncover subtle traces left by adversarial perturbations. Specifically, we analyze the high-frequency residual between similar images within the queries to detect adversarial examples. We evaluate our method against three advanced black-box attack strategies. Extensive experiments demonstrate that our approach achieves highly effective attack detection while significantly reducing FPR.

The static and homogeneous nature of traditional networks presents a significant challenge for our defense efforts. These characteristics enable an experienced attacker to quickly determine our network topology and gather detailed information about the internal hosts through systematic scanning techniques. Implementing a virtual network view can mitigate this by simulating a virtual topology, thereby consuming the attacker’s resources and time. However, deploying a virtual network view reduces network throughput and increase latency. Additionally, an improperly configured virtual network view can waste resources and degrade Quality of Service (QoS). Most existing studies have focused solely on the defender’s perspective, resulting in overly idealistic solutions that are ineffective in real-world scenarios. To address this, we propose VN-GT, a game-theoretic based model that optimizes virtual network deployment by considering both attackers and defenders. We provide a detailed example scenario, analyze the game’s equilibrium, and validate the effectiveness of our method through a real attack and defense experiment.

The rapid advancement of generative models has led to the widespread emergence of highly realistic synthetic images, making the detection of AI-generated content increasingly critical. In particular, diffusion models have recently achieved unprecedented levels of visual fidelity, further raising concerns. While most existing approaches rely on supervised learning, zero-shot detection methods have attracted growing interest due to their ability to bypass data collection and maintenance. Nevertheless, the performance of current zero-shot methods remains limited. In this paper, we introduce a novel zero-shot AI-generated image detection method. Unlike previous works that primarily focus on identifying artifacts in the final generated images, our work explores features within the image generation process that can be leveraged for detection. Specifically, we simulate the image sampling process via diffusion-based inversion and observe that the denoising outputs of generated images converge to the target image more rapidly than those of real images. Inspired by this observation, we compute the similarity between the original image and the outputs along the denoising trajectory, which is then used as an indicator of image authenticity.Since our method requires no training on any generated images, it avoids overfitting to specific generative models or dataset biases. Experiments across a wide range of generators demonstrate that our method achieves significant improvements over state-of-the-art supervised and zero-shot counterparts.

Pre-trained language models are increasingly being used in multi-document summarization tasks. However, these models need large-scale corpora for pre-training and are domain-dependent. Other non-neural unsupervised summarization approaches mostly rely on key sentence extraction, which can lead to information loss. To address these challenges, we propose a lightweight yet effective unsupervised approach called GLIMMER: a Graph and LexIcal features based unsupervised Multi-docuMEnt summaRization approach. It first constructs a sentence graph from the source documents, then automatically identifies semantic clusters by mining low-level features from raw texts, thereby improving intra-cluster correlation and the fluency of generated sentences. Finally, it summarizes clusters into natural sentences. Experiments conducted on Multi-News, Multi-XScience and DUC-2004 demonstrate that our approach outperforms existing unsupervised approaches. Furthermore, it surpasses state-of-the-art pre-trained multi-document summarization models (e.g. PEGASUS and PRIMERA) under zero-shot settings in terms of ROUGE scores. Additionally, human evaluations indicate that summaries generated by GLIMMER achieve high readability and informativeness scores. Our code is available at https://github.com/Oswald1997/GLIMMER.

Detection of face forgery videos remains a formidable challenge in the field of digital forensics, especially the generalization to unseen datasets and common perturbations. In this paper, we tackle this issue by leveraging the synergy between audio and visual speech elements, embarking on a novel approach through audio-visual speech representation learning. Our work is motivated by the finding that audio signals, enriched with speech content, can provide precise information effectively reflecting facial movements. To this end, we first learn precise audio-visual speech representations on real videos via a self-supervised masked prediction task, which encodes both local and global semantic information simultaneously. Then, the derived model is directly transferred to the forgery detection task. Extensive experiments demonstrate that our method outperforms the state-of-the-art methods in terms of cross-dataset generalization and robustness, without the participation of any fake video in model training.

Binary function matching compares two pieces of binary functions to identify their similarities, which has wide applications in the field of malware origin tracing, vulnerability searching, binary level plagiarism detection, etc. Up-to-date methods commonly independently map each function to an embedding and rarely consider fine-grained pairwise semantic similarity, which influences the accuracy of matching. Moreover, few methods are available to detect similarities between versions spanning a long period for cross-version vulnerability detection or patch positioning. To solve these issues, we propose a novel binary function matching method, which takes a pair of binary functions as input, and then computes a similarity score jointly on the pair through a specifical dual pairwise cross-attention network. Specially, we apply our method to detecting similarities between cross-version binaries. The experimental analysis demonstrates that FUNC-ESIM achieves promising results on the cross-version binary matching task, where the average recall@1 reaches 85.98%.

In order to detect malicious web scripts automatically, many detection methods using static features and machine learning are proposed. However, the existing detection methods can only detect web scripts of specific programming languages. This paper proposes the unified text features and abstract syntax tree(AST) node sequence features algorithm(UTANSA) that exploits the text feature classification method and AST node classification method, together with the corresponding unified method to enhance the generalization ability of the model. Through the algorithm, two unified approaches are proposed based on text features and AST node features respectively, so that the detection model can detect multi-language web scripts. We choose scripts written in the JavaScript(JS) and PHP languages for experimentation to evaluate our approach. The results show that the detection model trained with the proposed method has a similar detection effect as trained with only JS samples or PHP samples.

In recent years, scene text detection has experienced rapid development. Regression-based methods are currently a mainstream method for scene text detection, and the effect of bounding box regression is a major factor limiting their detection performance. The regression of bounding boxes is greatly affected by the aspect ratio of texts since the text in natural scenes varies greatly in height and width. However, the existing methods ignore the difference between the height and width of the text in the bounding box regression, which leads to an imperfect regression effect and thus suppresses the performance of the scene text detection. In this paper, we propose an Adaptive Smooth L1 Loss function (abbreviated as ASLL) for bounding box regression, which can adaptively determine the weight of each regression variable according to the current state of the model during the training process, so as to guide the bounding box to regress in a more critical direction. The experimental results demonstrate that ASLL achieves promising performance on scene text detection. Specially, an F-measure of 84.56% is achieved on CTW-1500 dataset, surpassing the state-of-the-art detectors, and the detection results on TotalText and ICDAR2015 datasets are competitive to those of state-of-the-art methods.

Abstract Reading text in images automatically has become an attractive research topic in computer vision. Specifically, end-to-end spotting of scene text has attracted significant research attention, and relatively ideal accuracy has been achieved on several datasets. However, most of the existing works overlooked the semantic connection between the scene text instances, and had limitations in situations such as occlusion, blurring, and unseen characters, which result in some semantic information lost in the text regions. The relevance between texts generally lies in the scene images. From the perspective of cognitive psychology, humans often combine the nearby easy-to-recognize texts to infer the unidentifiable text. In this paper, we propose a novel graph-based method for intermediate semantic features enhancement, called Text Relation Networks. Specifically, we model the co-occurrence relationship of scene texts as a graph. The nodes in the graph represent the text instances in a scene image, and the corresponding semantic features are defined as representations of the nodes. The relative positions between text instances are measured as the weights of edges in the established graph. Then, a convolution operation is performed on the graph to aggregate semantic information and enhance the intermediate features corresponding to text instances. We evaluate the proposed method through comprehensive experiments on several mainstream benchmarks, and get highly competitive results. For example, on the , our method surpasses the previous top works by 2.1% on the word spotting task.

Abuse of face swap techniques poses serious threats to the integrity and authenticity of digital visual media. More alarmingly, fake images or videos created by deep learning technologies, also known as Deepfakes, are more realistic, high-quality, and reveal few tampering traces, which attracts great attention in digital multimedia forensics research. To address those threats imposed by Deepfakes, previous work attempted to classify real and fake faces by discriminative visual features, which is subjected to various objective conditions such as the angle or posture of a face. Differently, some research devises deep neural networks to discriminate Deepfakes at the microscopic-level semantics of images, which achieves promising results. Nevertheless, such methods show limited success as encountering unseen Deepfakes created with different methods from the training sets. Therefore, we propose a novel Deepfake detection system, named FakeFilter, in which we formulate the challenge of unseen Deepfake detection into a problem of cross-distribution data classification, and address the issue with a strategy of domain adaptation. By mapping different distributions of Deepfakes into similar features in a certain space, the detection system achieves comparable performance on both seen and unseen Deepfakes. Further evaluation and comparison results indicate that the challenge has been successfully addressed by FakeFilter.

Confidential information firewalling with text classifier is to identify the text containing confidential information whose publication might be harmful to national security, business trade, or personal life. Traditional methods, e.g., listing a set of suspicious keywords together with regular-expression based filter, fail to solve the multi-topic phenomenon, i.e., one text containing the confidential information with different topics. In this paper, we propose a semi-supervised method, CIDetector, for multi-topic confidential information detection. We introduce coarse confidential polarity as prior knowledge into word embeddings, which can regularize the distribution of words to have a clear task classification boundary. Then we introduce a multi-attention network classifier to extract task-related features and model dependencies between features for multi-topic classification. Experiments are conducted by real-world data from WikiLeaks and demonstrated the superiority of our proposed method.

Confidential information firewalling with text classifiers is to recognize the text containing confidential information whose publication might pose a threat to national security, business trade, or personal life. Word embedding is a component of the detector and plays an important role. Existing word embeddings, e.g., Word2Vec, fail to learn a clear task classification boundary, i.e., the confidential polarities of words are opposite but the embedding vectors of the words are close to each other. We propose a confidentiality-oriented word embedding, CES2Vec, for confidential information detection. We embed confidentiality into semantics to catch both of them together, which can learn the word embedding with a clear task classification boundary. We use real-world data from WikiLeaks and conduct the comparison experiments of our CES2Vec and popular methods. The experimental results show that our proposed method is better than the previously reported methods in detecting confidential information.

Android malware family classification is an advanced task in Android malware analysis, detection and forensics. Existing methods and models have achieved a certain success for Android malware detection, but the accuracy and the efficiency are still not up to the expectation, especially in the context of multiple class classification with imbalanced training data. To address those challenges, we propose an Android malware family classification model by analyzing the code's specific semantic information based on sensitive opcode sequence. In this work, we construct a sensitive semantic feature-sensitive opcode sequence using opcodes, sensitive APIs, STRs and actions, and propose to analyze the code's specific semantic information, generate a semantic related vector for Android malware family classification based on this feature. Besides, aiming at the families with minority, we adopt an oversampling technique based on the sensitive opcode sequence. Finally, we evaluate our method on Drebin dataset, and select the top 40 malware families for experiments. The experimental results show that the Total Accuracy and Average AUC (Area Under Curve, AUC) reach 99.50% and 98.86% with 45. 17s per Android malware, and even if the number of malware families increases, these results remain good.

In Android malware detection, sensitive data-flows provide more accurate information on the application's behavior than regular features such as signatures and permissions. Currently, Android static taint analysis is widely adopted to identify sensitive data-flows because of its high code coverage and low false negative rate. However, existing static taint analysis tools cannot effectively analyze applications that adopt Android reflection mechanism. Reflection mechanism can block the control-flows and data-flows of the application. When constructing a call graph, the call information will point directly to the system's reflection processing method, rather than the actual method invoked by the application. This significantly affects the accurate representation of the application's behavior. To address this issue, this paper proposes a unified Android malware detection model based on reflective calls named URefFlow, in which the reflective call statement is replaced by the non-reflective call statement to make the reflective calls explicit by combining the parameters of the reflective calls into standard function calls. After extracting the complete sensitive data-flows with reflective calls from an application, we analyze the characteristics of these data-flows to determine whether the application is malicious. Evaluation results on thousands of applications show that URefFlow can achieve an impressive detection accuracy of 95.6% with a false positive rate of 0.8%. In addition, the proposed approach complements well with existing static stain analysis techniques.

In regard to relational databases, studies in this area typically focus on individual privacy leakage in one table. However, in reality, a database usually has many tables, some of them contain correlation information about individual, which can provide additional implication as background knowledge to attacker. In this paper, we innovatively propose a new method named MPP (Multi-table Privacy Preservation) which combines Lossy-join with Bucketization to enhance the individual privacy in database. We consider the privacy disclosure problem from the global sight of the entire dataset instead of a table. Based on this method, we not only solve the correlation information leakage by other tables, but also improve the data utility. Extensive experiments on 32.8GB real-world Express data demonstrate the effectiveness and efficiency of our approach in terms of data utility and computational cost.

PowerShell is so extremely powerful that we have seen that attackers are increasingly using PowerShell in their attack methods lately. In most cases, PowerShell malware arrives via spam email, using a combination of Microsoft Word documents to infect victims with its deadly payload. Nowadays, the de-obfuscation and analysis of PowerShell are still based on the manual analysis. However, as the number of malicious samples and obfuscation methods growing quickly, it is so slow that can't satisfy the demand. In this paper, we propose a de-obfuscation method of PowerShell called PSDEM which has two layers de-obfuscation to get original PowerShell scripts. One is extracting PowerShell scripts from much obfuscated document code. The other is de-obfuscating scripts including encoding, strings manipulation and code logic obfuscation. Meanwhile, we design an automatic de-obfuscation and analysis tool for malicious PowerShell scripts in Word documents based on PSDEM. We test the performance of the tool from the accuracy of de-obfuscation and the efficiency of time, and evaluation results show that it has a satisfactory performance. PSDEM improves the efficiency and accuracy rate for analyzing malicious PowerShell Scripts in Word documents, as well as provides a path in which further analysis for security experts to get more information about attacks.

Risk Score (RS) on Android is aiming at offering measurement to users for evaluating the apps' trustworthiness. Much work has been done to assess Android app's risk, but few jobs use various assessment systems to analyze Android apps with various malicious acts. However, it is hard for a single system to analyze those multiple categories Android apps. To overcome such limitations, we propose a multi-act classification model MRDroid for Android malware risk assessment in this paper, which presorts an app to one category, then uses the most suitable subsystem corresponding to that category to analyze the app for giving a RS. Base on this model, we implement an Android malware risk assessment system utilizing a machine learning solution with k-means algorithm for clustering benign and malware samples to various categories and the supervised algorithms for generating specific subsystems. It can be also used for Android malware detection under the condition of human confirmation. Experiments show that MRDroid provides high detection precision and offers stable and reliable risk assessment. Though testing our system using the dataset different from the system used, the result indicates it is also effective in detecting some unknown samples.

In this paper, we propose a model of email worm propagation based on a novel modeling method, Stochastic Game Nets (SGN), and use this model to analyze the rampant propagation issues of email worm. Combination the conceptual framework of SGN with the practical problems, we get some remarks based on the definitions of SGN in order to precisely describe the details of email worm propagation. An algorithm for solving the equilibrium strategy is presented to calculate the model of SGN. Finally we analyze our research result with several figures, such as infection rate and average propagation time. The results of our work can also offer some references for email users to defend email worms.