The proliferation of online misinformation has underscored the urgent need for robust, comprehensive verification tools. Existing approaches often lack the capacity for multi-modal analysis and evidence-based explainability, limiting their effectiveness in diverse, real-world scenarios. DeepVerify presents a complete end-to-end software solution for evidence-based online information verification, integrating previously published modules from ACMMM 2025 and supporting multi-lingual users. The system introduces a novel algorithm for detecting AI-generated content in images and text, extending prior capabilities that focused solely on video. The core idea of the algorithm is to compare the similarity of the input with information available online, allowing for the detection of AI-generated content. As a result, the system not only determines whether the data is created by humans or machines, but also provides specific evidence to support this decision. Unlike previous approaches that merely output a confidence score, this method offers concrete supporting evidence for its verdict, enhancing transparency and reliability in the verification process. Experiments conducted on benchmark datasets demonstrate that DeepVerify consistently outperforms current state-of-the-art methods, offering enhanced accuracy and transparent reasoning across multiple modalities. This work sets a new standard for explainable, reliable online information verification suitable for global audiences.

Worldwide image geolocalization, which aims to predict the GPS coordinates of any image on Earth, remains challenging due to global visual diversity. Recent generative approaches based on Retrieval-Augmented Generation (RAG) and Large Multimodal Models (LMMs) leverage candidates retrieved from fixed databases for reasoning, but often struggle with scenes that are absent from the reference set. In this work, we propose GeoSearch, an open-world geolocation framework that integrates web-scale reverse image search into the RAG pipeline. GeoSearch augments LMM prompts with database-retrieved coordinates and textual evidence extracted from web pages. To mitigate noise from irrelevant content, we introduce a two-layer filtering mechanism consisting of image matching, followed by confidence-based gating. Experiments on standard benchmarks Im2GPS3k and YFCC4k demonstrate the superiority of GeoSearch under leakage-aware evaluation. Our code and data are publicly available to support reproducibility.

The proliferation of multimedia content on social media has transformed how information is produced and consumed, enabling real-time coverage of global events but also accelerating the spread of misinformation, particularly during crises such as wars, natural disasters, and elections. The rise of synthetic media and the reuse of authentic content in misleading contexts further underscore the need for robust verification tools. In this paper, we present a comprehensive system developed for the ACM Multimedia 2025 Grand Challenge on Multimedia Verification. Our system evaluates both the authenticity and contextual accuracy of multimedia content in multilingual settings, producing expert-oriented verification reports alongside accessible summaries for the public. We propose a unified verification pipeline that integrates visual forensics, textual analysis, and multimodal reasoning, with a hybrid approach to detecting out-of-context (OOC) media via semantic similarity, temporal alignment, and geolocation cues. Extensive evaluations on the challenge benchmark demonstrate the system's effectiveness across diverse real-world scenarios. Our contributions advance the state of the art in multimedia verification while providing practical tools for journalists, fact-checkers, and researchers addressing information integrity in the digital age.

Large language models have emerged as a significant phenomenon due to their ability to produce natural text across various applications.However, the proliferation of generated text raises concerns regarding its potential misuse in fraudulent activities such as academic dishonesty, spam dissemination, and misinformation propagation.Prior studies have detected the generation of non-analogous text, which manifests numerous differences between original and generated text.We have observed that the similarity between the original text and its generation is notably higher than that between the generated text and its subsequent regeneration.To address this, we propose a novel approach named SimLLM, aimed at estimating the similarity between an input sentence and its generated counterpart to detect analogous machine-generated sentences that closely mimic human-written ones.Our empirical analysis demonstrates SimLLM's superior performance compared to existing methods.

Large Language Models (LLMs) have been garnering significant attention of AI researchers, especially following the widespread popularity of ChatGPT. However, due to LLMs' intricate architecture and vast parameters, several concerns and challenges regarding their quality assurance require to be addressed. In this paper, a fine-tuned GPT-based sentiment analysis model is first constructed and studied as the reference in AI quality analysis. Then, the quality analysis related to data adequacy is implemented, including employing the content-based approach to generate reasonable adversarial review comments as the wrongly-annotated data, and developing surprise adequacy (SA)-based techniques to detect these abnormal data. Experi-ments based on Amazon.com review data and a fine-tuned GPT model were implemented. Results were thoroughly discussed from the perspective of AI quality assurance to present the quality analysis of an LLM model on generated adversarial textual data and the effectiveness of using SA on anomaly detection in data quality assurance.

Adversarial images have been proposed to deceive deep neural networks (DNNs) by adding perturbations to the pixels. Unlike existing attacks, Su et al. [1] analyzed an attack in an extremely limited constraint where only one pixel was modified. However, their one-pixel attack is easy to recognize by humans. In this paper, we improve the attack to enable the deceit of both DNNs and humans. We conducted a human recognition analysis to prove our attack's effect. We then propose detection and defense methods against the attack by re-attacking the adversarial images. Our experimental results on the six most recent convolutional neural networks show that while our attack achieved approximately the same success rates and confidence scores as in the existing attack, our attack achieves a higher success rate for deceiving humans. Only 49.41 % of participants can recognize our attack even though 81.04 % participants have recognized the existing attack. OPA2D detects 99.33% of the existing attack and 100% of our attack and defends 92.00% of the existing attack and 95.33 % of our attack.

Hoang-Quoc Nguyen-Son, Tran Thao, Seira Hidano, Ishita Gupta, Shinsaku Kiyomoto. Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2021.

A visual homograph attack is a way that the attacker deceives the web users about which domain they are visiting by exploiting forged domains that look similar to the genuine domains. T. Thao et al. (IFIP SEC'19) proposed a homograph classification by applying conventional supervised learning algorithms on the features extracted from a single-character-based Structural Similarity Index (SSIM). This paper aims to improve the classification accuracy by combining their SSIM features with 199 features extracted from a N-gram model and applying advanced ensemble learning algorithms. The experimental result showed that our proposed method could enhance even 1.81% of accuracy and reduce 2.15% of false-positive rate. Furthermore, existing work applied machine learning on some features without being able to explain why applying it can improve the accuracy. Even though the accuracy could be improved, understanding the ground-truth is also crucial. Therefore, in this paper, we conducted an error empirical analysis and could obtain several findings behind our proposed approach.

Machine-translated text plays a crucial role in the communication of people using different languages. However, adversaries can use such text for malicious purposes such as plagiarism and fake review. The existing methods detected a machine-translated text only using the text’s intrinsic content, but they are unsuitable for classifying the machine-translated and human-written texts with the same meanings. We have proposed a method to extract features used to distinguish machine/human text based on the similarity between the intrinsic text and its back-translation. The evaluation of detecting translated sentences with French shows that our method achieves 75.0% of both accuracy and F-score. It outperforms the existing methods whose the best accuracy is 62.8% and the F-score is 62.7%. The proposed method even detects more efficiently the back-translated text with 83.4% of accuracy, which is higher than 66.7% of the best previous accuracy. We also achieve similar results not only with F-score but also with similar experiments related to Japanese. Moreover, we prove that our detector can recognize both machine-translated and machine-back-translated texts without the language information which is used to generate these machine texts. It demonstrates the persistence of our method in various applications in both low- and rich-resource languages.

Discriminating between computer-generated images (CGIs) and photographic images (PIs) is not a new problem in digital image forensics. However, with advances in rendering techniques supported by strong hardware and in generative adversarial networks, CGIs are becoming indistinguishable from PIs in both human and computer perception. This means that malicious actors can use CGIs for spoofing facial authentication systems, impersonating other people, and creating fake news to be spread on social networks. The methods developed for discriminating between CGIs and PIs quickly become outdated and must be regularly enhanced to be able to reduce these attack surfaces. Leveraging recent advances in deep convolutional networks, we have built a modular CGI--PI discriminator with a customized VGG-19 network as the feature extractor, statistical convolutional neural networks as the feature transformers, and a discriminator. We also devised a probabilistic patch aggregation strategy to deal with high-resolution images. This proposed method outperformed a state-of-the-art method and achieved accuracy up to 100%.

Making computer-generated (CG) images more difficult to detect is an interesting problem in computer graphics and security. While most approaches focus on the image rendering phase, this paper presents a method based on increasing the naturalness of CG facial images from the perspective of spoofing detectors. The proposed method is implemented using a convolutional neural network (CNN) comprising two autoencoders and a transformer and is trained using a black-box discriminator without gradient information. Over 50% of the transformed CG images were not detected by three state-of-the-art spoofing detectors. This capability raises an alarm regarding the reliability of facial authentication systems, which are becoming widely used in daily life.

Time-related information in message posted on-line is one type of sensitive information targeted by attackers, one reason that sharing information online can be risky. Therefore, time information should be anonymized before it is posted in online social networks (OSNs). One approach to reducing the risk is to anonymize the personal information by removing temporal phrases, but this makes the anonymous message loses too much information. We have proposed a system for creating anonymous fingerprints about temporal phrases to cover most of potential cases of OSN disclosure. The fingerprints not only anonymize time-related information but also can be used to identify a person who has disclosed information about the user. In experiment with 16,647 different temporal phrases extracted from about 16 million tweets, the average number of fingerprints created for an OSN message is 526.05 fingerprints. This is significantly better than the 409.58 fingerprints of the state-of-the-art previous detection temporal phrases algorithm. Fingerprints are quantified using a modified normalized certainty penalty metric to ensure that an appropriate level of information anonymity is used for each user's friend. The algorithm works well not only for temporal phrases in message posted on social networks but also for other types of phrases (such as location and objective ones) or other areas (religion, politics, military, etc.).

One approach to overcoming the problem of too much information about a user being disclosed on social networking services (by the user or by the user's friends) through natural language texts (blogs, comments, status updates, etc.) is to anonymize the texts. However, determining which information is sensitive and should thus be anonymized is a challenging problem. Sensitive information is any information about a user that could be used to identify the user. We have developed an algorithm that anonymizes sensitive information in text to be posted by generalization. Synonyms for the anonymized information are used as fingerprints for detecting a discloser of the information. The fingerprints are quantified using the modified discernability metric to enable an appropriate level of anonymity to be used for each group of the user's friends. The fingerprints cannot be converted into another one to incorrectly identify a person who has revealed sensitive information. Use of the algorithm to control the disclosure of information on Facebook demonstrated that it works well not only in social networking but also in other areas (health, religion, politics, military, etc.) that store sensitive information.