What does coordinated inauthentic behavior look like on TikTok?

We introduce a new framework for detecting coordination in video-first platforms, uncovering influence campaigns using synthetic voices, split-screen tactics, and cross-account duplication.
📄https://arxiv.org/abs/2505.10867

We constantly ask our apps where to visit, eat or drink.
AI tells us, and most of the time, we follow it. The loop continues.
But do AIs favor certain places? How would we even know if we don’t own the platforms?
We modeled this complex phenomenon, and results are fascinating!
Spoiler: rich get…

IU's Observatory on Social Media defends citizens from online manipulation – the opposite of censorship
osome.iu.edu/research/blo...

Preprint is available on arXiv arxiv.org/abs/2412.14663

This work would not have been possible without the other amazing coauthors @luceriluc.bsky.social @frafabbri.bsky.social @emilioferrara.bsky.social

Bonus Pic: myself beyond excited to stand next to my poster!

Our work provides a scalable approach for online moderation teams, public institutions, and independent organizations to audit the health of online environments—especially crucial during political events such as election cycles.

2. We explore how our multimodal framework exhibits foundation model behavior in detecting online information operations. Our results show that pretraining IOHunter on past IO datasets enables it to generalize to new, emerging IOs with only a few labeled examples for fine-tuning.

Key takeaways:

1. We propose a multimodal framework that effectively integrates textual and graph information using a cross-attention mechanism, which is then processed by a GNN.

Can we effectively detect covert Information Operations (IOs) that attempt to manipulate socio-political debates on social media?

This is the focus of our work, "IOHunter: Graph Foundation Model to Uncover Online Information Operations", just presented at the #AAAI #AAAI2025

The three horsemen of social media: brain rot, anxiety and foreign interference
voxeurop.eu/en/social-me...

Read our preprint available on arXiv at: arxiv.org/abs/2412.14663

Our effort highlights the critical role of multi-modality in modeling malicious user behavior, the value of attention to weight the modalities, and how we can advance toward a GFM for the IO Detection task by pre-training our architecture on a dataset of previous IOs.

Our work demonstrates how a multi-modal framework based on GNN+LM and massive pre-training produces a model that effectively generalizes to IOs not present in the original training dataset — the most realistic scenario for IO detection.

Our model delivers substantial improvements over current IO detection methods across three learning tasks:

1️⃣ Supervised IO Detection
2️⃣ Scarcely-Labeled Supervised IO Detection
3️⃣ Cross-IO Detection (with minimal or no labeled data from emerging IOs)

Maintaining the integrity of online discourse is essential for safeguarding fair democratic processes.

Our multi-modal learning framework IOHunter integrates both content and contextual information to identify actors attempting to manipulate online discussions - i.e., IO Drivers

"IOHunter: Graph Foundation Model to Uncover Online Information Operations" goes to AAAI'25!
This is the result of an incredible collaboration with @luceriluc.bsky.social @frafabbri.bsky.social and @emilioferrara.bsky.social

Read the entire thread for a summary and the link to the preprint.

New evidence of cross-platform foreign interference on social media during the 2024 U.S. Election that drove the spread of highly-partisan, low-credibility, and conspiratorial content, from Cinus, Minici, @luceriluc.bsky.social @emilioferrara.bsky.social arxiv.org/pdf/2410.22716