A paper co-authored by researchers from SALVUS partner, the University of Criminal Investigation and Police Studies (UCIPS), in Belgrade, has been accepted at TELFOR 2026, the 34th Telecommunications Forum. Automatic Detection of Predatory Online Conversations using Small-scale LLMs asks whether compact, openly available AI language models can learn to recognise predatory conversations using modest, local computing hardware. The study was partially conducted within SALVUS.
The authors are Aleksandar Miljković, Darko Stefanović and Srđan Sladojević (Faculty of Technical Sciences, University of Novi Sad), Milan Gnjatović and Zvonimir Ivanović (University of Criminal Investigation and Police Studies), and Jordan Atanasijević and Slaviša Ilić (Applied Mathematics and Electronics Centre, Belgrade).
Why detection is hard
Online grooming is gradual. Offenders build trust, emotional dependence and control step by step, and they adapt their language to get past platform safety filters, so keyword or rule-based detection falls short. Real grooming conversations are also rarely available for research, for ethical and legal reasons.
Large AI models can read context, but adapting them usually requires heavy computing infrastructure. The research team wanted to know whether smaller models, adapted on an ordinary workstation, could do the job.
What the team did
- Five small open models from different developers, from 3 to 7 billion parameters: Qwen3-4B, Qwen2.5-7B, Phi-4-mini, Mistral-7B and Llama-3.2-3B.
- Two rounds of testing: first as released, then after parameter-efficient fine-tuning (LoRA and its quantised variant, QLoRA), which trains only a small set of added parameters.
- An established research dataset, VTPAN, built for early sexual predator detection. Its predatory conversations come from decoy operations in which adults posed as children. Models were trained on 6,588 conversation segments and tested on 15,329, of which 1,466 were predatory.
- One laptop: all fine-tuning ran on a single Apple MacBook Pro (M5 chip, 48 GB memory).
What they found
- Out of the box, no model was reliable. Balanced accuracy ranged from 0.49 to 0.60, close to chance.
- After fine-tuning, every model performed strongly. Balanced accuracy rose to 0.98–0.99. For predatory conversations specifically, F1 scores (which weigh missed cases against false alarms; 1.0 is perfect) rose from 0.10–0.22 to 0.93–0.98.
- False alarms fell sharply. Qwen2.5-7B, for example, wrongly flagged 6,230 harmless conversations before fine-tuning and 76 after.
- Bigger was not automatically better. Before fine-tuning, all three compact models (3–4B parameters) achieved higher balanced accuracy than the larger Mistral-7B. The authors link this to advances in newer model families rather than raw size.
Why it matters for SALVUS
SALVUS works to make online and undercover child sexual abuse investigations more effective and legally sound. Investigators face large volumes of chat data, and this study suggests that capable triage models do not have to depend on large computing infrastructure: they can be adapted and run on hardware a research team or unit may already have.
As SALVUS research on child-centred investigations has found, technology must support, not replace, human judgement and child-rights sensitive practice. Tools like these can help prioritise what investigators review; decisions stay with trained professionals.
Limitations the authors note
- The data is in English only. Other languages remain untested, partly because comparable annotated datasets are scarce.
- The predatory conversations involve adult decoys posing as children, which may not fully reflect interactions with real child victims.
- The models classified whole conversation segments. Early detection, while a conversation is still under way, was not tested.
See it at TELFOR 2026
TELFOR takes place on 24–26 November 2026 at the Crowne Plaza and the Palace of Science in Belgrade, Serbia, with online participation available. It is organised by the Telecommunications Society and the School of Electrical Engineering of the University of Belgrade, with IEEE Region 8 and the IEEE Serbia and Montenegro Section. Visit the TELFOR website.
This study was partially conducted within the SALVUS project (Grant Agreement No. 101225758) and partially within research project AMEC/01/25-27 of the Applied Mathematics and Electronics Centre.
READ FULL ARTICLE
ℹ️ SALVUS Project – Funding acknowledgement
▪️ Funded by the European Union. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or the European Research Executive Agency. Neither the European Union nor the granting authority can be held responsible for them.
▪️Project funded by Swiss Confederation. State Secretariat for Education Research and Innovation SERI.

