LLMs Have Rhythm: Fingerprinting Large Language Models Using Inter-Token Times and Network Traffic Analysis
As Large Language Models (LLMs) become increasingly integrated into many technological ecosystems across various domains and industries, identifying which model is deployed or being interacted with is critical for the security and trustworthiness of the systems. Current verification methods typicall...
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2025-01-01
|
| Series: | IEEE Open Journal of the Communications Society |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/11026013/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1849320453650251776 |
|---|---|
| author | Saeif Alhazbi Ahmed Hussain Gabriele Oligeri Panos Papadimitratos |
| author_facet | Saeif Alhazbi Ahmed Hussain Gabriele Oligeri Panos Papadimitratos |
| author_sort | Saeif Alhazbi |
| collection | DOAJ |
| description | As Large Language Models (LLMs) become increasingly integrated into many technological ecosystems across various domains and industries, identifying which model is deployed or being interacted with is critical for the security and trustworthiness of the systems. Current verification methods typically rely on analyzing the generated output to determine the source model. However, these techniques are susceptible to adversarial attacks, operate in a post-hoc manner, and may require access to the model weights to inject a verifiable fingerprint. In this paper, we propose a novel passive fingerprinting framework that operates in real-time and remains effective even under encrypted network traffic conditions. Our method leverages the intrinsic autoregressive generation nature of language models, which generate text one token at a time based on all previously generated tokens, creating a unique temporal pattern-like a rhythm or heartbeat-that persists even when the output is streamed over a network. We find that measuring the Inter-Token Times (ITTs)–time intervals between consecutive tokens-can identify different language models with high accuracy. We develop a Deep Learning (DL) pipeline to capture these timing patterns using network traffic analysis and evaluate it on 16 Small Language Models (SLMs) and 10 proprietary LLMs across different deployment scenarios, including local host machine (GPU/CPU), Local Area Network (LAN), Remote Network, and when using Virtual Private Network (VPN). Our experimental results demonstrate high classification performance with weighted F1-scores of 85% when tested on a different day, 74% across different networks, and 71% when traffic is tunneled through a VPN connection. This work opens a new avenue for model identification in real-world scenarios and contributes to more secure and trustworthy language model deployment. |
| format | Article |
| id | doaj-art-d251fac42a81458ab60c0e5ffe450053 |
| institution | Kabale University |
| issn | 2644-125X |
| language | English |
| publishDate | 2025-01-01 |
| publisher | IEEE |
| record_format | Article |
| series | IEEE Open Journal of the Communications Society |
| spelling | doaj-art-d251fac42a81458ab60c0e5ffe4500532025-08-20T03:50:06ZengIEEEIEEE Open Journal of the Communications Society2644-125X2025-01-0165050507110.1109/OJCOMS.2025.357701611026013LLMs Have Rhythm: Fingerprinting Large Language Models Using Inter-Token Times and Network Traffic AnalysisSaeif Alhazbi0https://orcid.org/0000-0002-7884-5025Ahmed Hussain1https://orcid.org/0000-0003-4732-9543Gabriele Oligeri2https://orcid.org/0000-0002-9637-0430Panos Papadimitratos3https://orcid.org/0000-0002-3267-5374College of Science and Engineering, Hamad Bin Khalifa University, Doha, QatarNetworked Systems Security Group, KTH Royal Institute of Technology, Stockholm, SwedenCollege of Science and Engineering, Hamad Bin Khalifa University, Doha, QatarNetworked Systems Security Group, KTH Royal Institute of Technology, Stockholm, SwedenAs Large Language Models (LLMs) become increasingly integrated into many technological ecosystems across various domains and industries, identifying which model is deployed or being interacted with is critical for the security and trustworthiness of the systems. Current verification methods typically rely on analyzing the generated output to determine the source model. However, these techniques are susceptible to adversarial attacks, operate in a post-hoc manner, and may require access to the model weights to inject a verifiable fingerprint. In this paper, we propose a novel passive fingerprinting framework that operates in real-time and remains effective even under encrypted network traffic conditions. Our method leverages the intrinsic autoregressive generation nature of language models, which generate text one token at a time based on all previously generated tokens, creating a unique temporal pattern-like a rhythm or heartbeat-that persists even when the output is streamed over a network. We find that measuring the Inter-Token Times (ITTs)–time intervals between consecutive tokens-can identify different language models with high accuracy. We develop a Deep Learning (DL) pipeline to capture these timing patterns using network traffic analysis and evaluate it on 16 Small Language Models (SLMs) and 10 proprietary LLMs across different deployment scenarios, including local host machine (GPU/CPU), Local Area Network (LAN), Remote Network, and when using Virtual Private Network (VPN). Our experimental results demonstrate high classification performance with weighted F1-scores of 85% when tested on a different day, 74% across different networks, and 71% when traffic is tunneled through a VPN connection. This work opens a new avenue for model identification in real-world scenarios and contributes to more secure and trustworthy language model deployment.https://ieeexplore.ieee.org/document/11026013/Large language modelssmall language modelsfingerprintingnetwork traffic analysisdeep learningnetwork security |
| spellingShingle | Saeif Alhazbi Ahmed Hussain Gabriele Oligeri Panos Papadimitratos LLMs Have Rhythm: Fingerprinting Large Language Models Using Inter-Token Times and Network Traffic Analysis IEEE Open Journal of the Communications Society Large language models small language models fingerprinting network traffic analysis deep learning network security |
| title | LLMs Have Rhythm: Fingerprinting Large Language Models Using Inter-Token Times and Network Traffic Analysis |
| title_full | LLMs Have Rhythm: Fingerprinting Large Language Models Using Inter-Token Times and Network Traffic Analysis |
| title_fullStr | LLMs Have Rhythm: Fingerprinting Large Language Models Using Inter-Token Times and Network Traffic Analysis |
| title_full_unstemmed | LLMs Have Rhythm: Fingerprinting Large Language Models Using Inter-Token Times and Network Traffic Analysis |
| title_short | LLMs Have Rhythm: Fingerprinting Large Language Models Using Inter-Token Times and Network Traffic Analysis |
| title_sort | llms have rhythm fingerprinting large language models using inter token times and network traffic analysis |
| topic | Large language models small language models fingerprinting network traffic analysis deep learning network security |
| url | https://ieeexplore.ieee.org/document/11026013/ |
| work_keys_str_mv | AT saeifalhazbi llmshaverhythmfingerprintinglargelanguagemodelsusingintertokentimesandnetworktrafficanalysis AT ahmedhussain llmshaverhythmfingerprintinglargelanguagemodelsusingintertokentimesandnetworktrafficanalysis AT gabrieleoligeri llmshaverhythmfingerprintinglargelanguagemodelsusingintertokentimesandnetworktrafficanalysis AT panospapadimitratos llmshaverhythmfingerprintinglargelanguagemodelsusingintertokentimesandnetworktrafficanalysis |