Malicious actors profiling and detection in Online Social Networks through Artificial Intelligence – MARTINI

W
We proposed Bi-LORA (Harnessing the Power of Large Vision Language Models for Synthetic Image Detection), a novel method based on Visual Language Models (VLMs).
o Bi-LORA leverages the generalization power of VLMs, trained on large-scale internet data, to offer strong zero-shot capabilities for feature extraction.

This work was accepted for presentation at the ICASSP 2024 conference
o Bi-LORA demonstrated its potential to significantly enhance multimodal feature extraction and AI-generated image detection, providing a robust framework for extracting generalizable features.

We are extending our work to develop a multimodal model capable of detecting manipulated content across images, videos, and audio.
The model aims to integrate cross-modal features to improve accuracy and reliability while addressing challenges like cross-modal consistency.

? Synthetic Image Detection Using Mixture of Knowledge Distillation from Vision-Language Models (Workshop on MultiMedia Forensics in the Wild, Kolkata, India, 2024.)
? FIDAVL: Fake Image Detection and Attribution Using Vision-Language Models (27th International Conference on Pattern Recognition, India, Kolkata, 2024.)
? On the Detection of AI-Generated Images and Videos (4th International Conference on Image Processing and Vision Engineering, 2024)
? Advances in AI-Generated Images and Videos (International Journal of Interactive Multimedia and Artificial Intelligence (IJIMAI), 2024)
? Bi-LORA: A Vision-Language Approach for Synthetic Image Detection (Expert Systems Journal, 2024)
? Generation and detection of manipulated multimodal audiovisual content: Advances, trends and open challenges (Information Fusion Journal, 2024)
? Harnessing the Power of Large Vision Language Models for Synthetic Image Detection (2024 IEEE International Conference on Acoustics, Speech, and Signal Processing, ICASSP, 2024)
? DeeCLIP: A Robust Transformer-Based Approach for AI-Generated Image Detection (IEEE/CVF Conference on Computer Vision and Pattern Recognition 2025), In process.
? On the Effectiveness of Vision Mamba for AI-Generated Image Detection (International Journal of Computer Vision, IJCV, 2024), In process.

There is no doubt that Online Social Networks have changed the way people communicate and interact. These platforms have helped to break down barriers and to ease worldwide communication, to share multimedia content or to easily know what is happening in almost any point in the world. Facebook (2004) or Twitter (2006) but also other social media platforms such as Youtube (2005) or messaging services such as WhatsApp (2009) have promoted this new era. However, the meteoric rise of the use of these platforms has also been followed by continuous attempts to subvert their original purposes, using them by the so-called malicious actors for evil purposes. Examples of this can be the generation of mistrust against vaccines, the creation of content supporting climate denial theories, or disinformation campaigns trying to manipulate people's opinion to alter the results of democratic elections. Therefore, there is a wide variety of different types of attempts to undermine social networks and to distorse public discourse.

In this project, we seek to reveal and detect the presence of malicious actors in Online Social Networks and to profile these actors through a multidisciplinary approach, including experts in the use of novel computational techniques from the Artificial Intelligence and Computer Vision fields and experts from the Behavioural Sciences field. The project will also leverage a multimodality approach, analysing text, images, videos, audio network structures, interactions between users and trust perceptions. From the Behavioural Sciences research field, different approaches, including psychology, discourse analysis, sociology and human-computer interaction will allow us to build a taxonomy of the different malicious actors and to profile them. From the Artificial Intelligence field, tools such as Deep Learning, advanced Natural Language Processing, and Social Network Analysis provide us with the necessary instruments to build a software tool for the analysis and detection of these malicious actors with novel features such as multilingualism, explainability, and a strong focus on an open-source solution.