Slide
Centre Interdisciplinaire
de Recherche et d’Innovation
en Cybersécurité et Société
de Recherche et d’Innovation
en Cybersécurité et Société
1.
Pimentel, A.; Zhu, Y.; Falk, T. H.
Partial Audio Deepfake Detection: Are We Really Detecting Synthetic Media or Just Dataset Content Biases? Article d'actes
Dans: IEEE Conf. Artif. Intell., CAI, p. 1846–1851, Institute of Electrical and Electronics Engineers Inc., 2026, ISBN: 979-833156039-3 (ISBN), (Journal Abbreviation: IEEE Conf. Artif. Intell., CAI).
Résumé | Liens | BibTeX | Étiquettes: Audio signal processing, Computational linguistics, Condition, Detection models, Information integrity, Language model, Large datasets, Learning models, Optimistics, Performance, Self-supervised learning, Speech communication, Speech models, Speech recognition, Speech signals, Synthetic media, Transcription
@inproceedings{pimentelPartialAudioDeepfake2026,
title = {Partial Audio Deepfake Detection: Are We Really Detecting Synthetic Media or Just Dataset Content Biases?},
author = {A. Pimentel and Y. Zhu and T. H. Falk},
url = {https://www.scopus.com/pages/publications/105042045476?origin=resultslist},
doi = {10.1109/CAI68641.2026.11536622},
isbn = {979-833156039-3 (ISBN)},
year = {2026},
date = {2026-01-01},
booktitle = {IEEE Conf. Artif. Intell., CAI},
pages = {1846–1851},
publisher = {Institute of Electrical and Electronics Engineers Inc.},
abstract = {In recent years, the generation of highly realistic audio deepfakes has raised significant concerns regarding privacy and information integrity. While most research has focused on fully bonafide or spoofed speech, partial deep-fakes, where only segments of an utterance are manipulated, remain less explored. Given the nature of the task, existing datasets are relying on large language models to manipulate bonafide speech signals into partial deepfakes by altering, deleting, or replacing segments with synthetic content. These manipulations may alter the semantics and sentiment of the generated content, creating biases that can be captured by deepfake detection models, leading to overly optimistic performance and poor generalizability. In this work, we investigate the presence of such biases in two popular partial deepfake datasets, namely AV-Deepfake1M and PartialEdit. We explore four self-supervised learning models, two relying only on text transcriptions (RoBERTa and RoBERTa-sentiment) and two relying on universal speech representations (WavLM Large and Wav2vec2-XLSR). Our experiments show that, while speech models achieve higher in-domain performance, they do not generalize to out-of-domain conditions. In turn, models trained on only transcribed speech can effectively distinguish manipulated content, achieving up to 97.8% AUC in-domain and nearly 70% AUC out-of-domain. These results suggest that linguistic patterns and dataset confounds may indeed be biasing partial deepfake detection models, leading to poor generalizability. © 2026 IEEE.},
note = {Journal Abbreviation: IEEE Conf. Artif. Intell., CAI},
keywords = {Audio signal processing, Computational linguistics, Condition, Detection models, Information integrity, Language model, Large datasets, Learning models, Optimistics, Performance, Self-supervised learning, Speech communication, Speech models, Speech recognition, Speech signals, Synthetic media, Transcription},
pubstate = {published},
tppubtype = {inproceedings}
}
In recent years, the generation of highly realistic audio deepfakes has raised significant concerns regarding privacy and information integrity. While most research has focused on fully bonafide or spoofed speech, partial deep-fakes, where only segments of an utterance are manipulated, remain less explored. Given the nature of the task, existing datasets are relying on large language models to manipulate bonafide speech signals into partial deepfakes by altering, deleting, or replacing segments with synthetic content. These manipulations may alter the semantics and sentiment of the generated content, creating biases that can be captured by deepfake detection models, leading to overly optimistic performance and poor generalizability. In this work, we investigate the presence of such biases in two popular partial deepfake datasets, namely AV-Deepfake1M and PartialEdit. We explore four self-supervised learning models, two relying only on text transcriptions (RoBERTa and RoBERTa-sentiment) and two relying on universal speech representations (WavLM Large and Wav2vec2-XLSR). Our experiments show that, while speech models achieve higher in-domain performance, they do not generalize to out-of-domain conditions. In turn, models trained on only transcribed speech can effectively distinguish manipulated content, achieving up to 97.8% AUC in-domain and nearly 70% AUC out-of-domain. These results suggest that linguistic patterns and dataset confounds may indeed be biasing partial deepfake detection models, leading to poor generalizability. © 2026 IEEE.
2.
Temmar, D. E.; Hamadene, A.; Nallaguntla, V.; Fursule, A.; Allili, M. S.; Kshirsagar, S.; Avila, A. R.
Phonetic Analysis of Real and Synthetic Speech Using HuBERT Embeddings: Perspectives for Deepfake Detection Article d'actes
Dans: Conf. Proc. IEEE Int. Conf. Syst. Man Cybern., p. 86–91, Institute of Electrical and Electronics Engineers Inc., 2025, ISBN: 1062922X (ISSN); 979-833153358-8 (ISBN), (Journal Abbreviation: Conf. Proc. IEEE Int. Conf. Syst. Man Cybern.).
Résumé | Liens | BibTeX | Étiquettes: Artificial intelligence, Audio acoustics, Audio DeepFake Detection, Audio signal processing, Embeddings, Hu-BERT, KL-divergence, Linguistics, Phoneme and word Embedding, Phonetic analysis, Security systems, Self-Supervised Speech Representation, Speech analysis, Speech communication, Speech processing, Speech synthesis, Synthetic speech, Text to speech, Voice conversion
@inproceedings{temmarPhoneticAnalysisReal2025,
title = {Phonetic Analysis of Real and Synthetic Speech Using HuBERT Embeddings: Perspectives for Deepfake Detection},
author = {D. E. Temmar and A. Hamadene and V. Nallaguntla and A. Fursule and M. S. Allili and S. Kshirsagar and A. R. Avila},
url = {https://www.scopus.com/pages/publications/105033145913?origin=resultslist},
doi = {10.1109/SMC58881.2025.11343334},
isbn = {1062922X (ISSN); 979-833153358-8 (ISBN)},
year = {2025},
date = {2025-01-01},
booktitle = {Conf. Proc. IEEE Int. Conf. Syst. Man Cybern.},
pages = {86–91},
publisher = {Institute of Electrical and Electronics Engineers Inc.},
abstract = {The growing sophistication of speech generated by Artificial Intelligence (AI) has introduced new challenges in audio deepfake detection. Text-to-speech (TTS) and voice conversion (VC) technologies can now produce convincing synthetic speech with high quality and intelligibility. This poses a serious threat to voice biometric security systems, such as automatic speaker recognition. It also increases the risks associated to the spread of spoken disinformation, where synthetic voices can be used to disseminate malicious content. In this study, we conduct an analysis of real and synthetic speech at phonetic and word levels. For that, a parallel dataset comprising real and synthetic speech signals were developed based on a subset of the LibriSpeech ASR corpus. Synthetic speech samples were generated using two TTS and one VC systems: Coqui TTS, VITS TTS, and StarGANv2 VC. We adopted HuBERT, a self-supervised speech model, to extract speech embeddings. The motivation for using this model stems from its ability to recognize sound units corresponding to the so-called pseudo phonemes. Our analysis is based on the KL divergence (KLD) between the distributions of synthetic and real phonemes, which allowed us to rank synthetic phonemes based on their alignment with their real counterpart. We also trained several classifiers per phoneme to distinguish between real and synthetic samples. We then compute the correlations between KLD and accuracies per phoneme. Besides showing a list of phonemes that are more discriminative, our findings suggest that vowels correlate better with the classifiers' performance, suggesting that the KLD can be an indicator of the most distinguishable phonemes for deepfake detection. © 2025 IEEE.},
note = {Journal Abbreviation: Conf. Proc. IEEE Int. Conf. Syst. Man Cybern.},
keywords = {Artificial intelligence, Audio acoustics, Audio DeepFake Detection, Audio signal processing, Embeddings, Hu-BERT, KL-divergence, Linguistics, Phoneme and word Embedding, Phonetic analysis, Security systems, Self-Supervised Speech Representation, Speech analysis, Speech communication, Speech processing, Speech synthesis, Synthetic speech, Text to speech, Voice conversion},
pubstate = {published},
tppubtype = {inproceedings}
}
The growing sophistication of speech generated by Artificial Intelligence (AI) has introduced new challenges in audio deepfake detection. Text-to-speech (TTS) and voice conversion (VC) technologies can now produce convincing synthetic speech with high quality and intelligibility. This poses a serious threat to voice biometric security systems, such as automatic speaker recognition. It also increases the risks associated to the spread of spoken disinformation, where synthetic voices can be used to disseminate malicious content. In this study, we conduct an analysis of real and synthetic speech at phonetic and word levels. For that, a parallel dataset comprising real and synthetic speech signals were developed based on a subset of the LibriSpeech ASR corpus. Synthetic speech samples were generated using two TTS and one VC systems: Coqui TTS, VITS TTS, and StarGANv2 VC. We adopted HuBERT, a self-supervised speech model, to extract speech embeddings. The motivation for using this model stems from its ability to recognize sound units corresponding to the so-called pseudo phonemes. Our analysis is based on the KL divergence (KLD) between the distributions of synthetic and real phonemes, which allowed us to rank synthetic phonemes based on their alignment with their real counterpart. We also trained several classifiers per phoneme to distinguish between real and synthetic samples. We then compute the correlations between KLD and accuracies per phoneme. Besides showing a list of phonemes that are more discriminative, our findings suggest that vowels correlate better with the classifiers' performance, suggesting that the KLD can be an indicator of the most distinguishable phonemes for deepfake detection. © 2025 IEEE.
3.
Jalleli, O.; Zhu, Y.; Falk, T. H.
Audio-Visual Cross-Attention for Improved Deepfake Video Detection and Forgery Localization Article d'actes
Dans: Conf. Proc. IEEE Int. Conf. Syst. Man Cybern., p. 75–79, Institute of Electrical and Electronics Engineers Inc., 2025, ISBN: 1062922X (ISSN); 979-833153358-8 (ISBN), (Journal Abbreviation: Conf. Proc. IEEE Int. Conf. Syst. Man Cybern.).
Résumé | Liens | BibTeX | Étiquettes: Artificial intelligence, Audio acoustics, Audio signal processing, Audio-visual, Computer vision, Crossmodal attention, Deepfake detection, Forgery, Generative AI, Localisation, Multi-modal, Video detection, Video forgeries, Visual modalities
@inproceedings{jalleliAudioVisualCrossAttentionImproved2025,
title = {Audio-Visual Cross-Attention for Improved Deepfake Video Detection and Forgery Localization},
author = {O. Jalleli and Y. Zhu and T. H. Falk},
url = {https://www.scopus.com/pages/publications/105033150623?origin=resultslist},
doi = {10.1109/SMC58881.2025.11342953},
isbn = {1062922X (ISSN); 979-833153358-8 (ISBN)},
year = {2025},
date = {2025-01-01},
booktitle = {Conf. Proc. IEEE Int. Conf. Syst. Man Cybern.},
pages = {75–79},
publisher = {Institute of Electrical and Electronics Engineers Inc.},
abstract = {With the emergence of multi-modal generative models, synthesized videos are becoming increasingly realistic, making the detection of deepfakes extremely challenging. While several video deepfake detection models have shown promising performance, their focus has been primarily on the visual modality. To overcome this limitation, we propose a dualstream framework that fuses visual and auditory information via cross-attention computed between embeddings extracted from pre-trained video and audio encoders. Additionally, we design a weakly-supervised forgery localization head that infers frame-level forgery scores from coarse segment-level labels, minimizing the need for fine-grained annotations and allowing for forgery location characterization. In this paper, we describe our preliminary results showing the proposed model outperforming state-of-the-art detectors on both frame-level localization and sequence-level deepfake detection tasks. Ongoing work focuses on investigating the complementarity between the visual and auditory modalities to improve model robustness and explainability. © 2025 IEEE.},
note = {Journal Abbreviation: Conf. Proc. IEEE Int. Conf. Syst. Man Cybern.},
keywords = {Artificial intelligence, Audio acoustics, Audio signal processing, Audio-visual, Computer vision, Crossmodal attention, Deepfake detection, Forgery, Generative AI, Localisation, Multi-modal, Video detection, Video forgeries, Visual modalities},
pubstate = {published},
tppubtype = {inproceedings}
}
With the emergence of multi-modal generative models, synthesized videos are becoming increasingly realistic, making the detection of deepfakes extremely challenging. While several video deepfake detection models have shown promising performance, their focus has been primarily on the visual modality. To overcome this limitation, we propose a dualstream framework that fuses visual and auditory information via cross-attention computed between embeddings extracted from pre-trained video and audio encoders. Additionally, we design a weakly-supervised forgery localization head that infers frame-level forgery scores from coarse segment-level labels, minimizing the need for fine-grained annotations and allowing for forgery location characterization. In this paper, we describe our preliminary results showing the proposed model outperforming state-of-the-art detectors on both frame-level localization and sequence-level deepfake detection tasks. Ongoing work focuses on investigating the complementarity between the visual and auditory modalities to improve model robustness and explainability. © 2025 IEEE.