Authors:
Joanna Gajewska, Alicja Martinek, and Ewelina Bartuzi-Trokielewicz
Where published:
Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) Workshops, 2026, pp. 251-260
Abstract:
The study evaluates how well current audio deepfake detection models perform in real-world conditions, particularly on recently observed fraud cases such as scam advertisements. The authors compiled a dataset of verified real-world deepfake audio and tested state-of-the-art detectors that were originally trained on standard benchmark datasets. They found that these models perform poorly on real-world samples, revealing a significant generalization gap. To investigate this, they experimented with modifying training data by adding realistic distortions like noise, compression, and signal degradation, which led to only modest improvements. However, incorporating actual incident-based (real-world) deepfake samples into training significantly improved detection performance. They also explored aggregation techniques that combine predictions across audio segments to better detect partially manipulated recordings. Overall, the study highlights that current detectors are not robust enough for real-world use unless they are trained and evaluated on more realistic, up-to-date data.
Dataset names (used for):
- Incident-Based Dataset: real-world fraudulent advertisement audio;
- IBD-partial: partial deepfake evaluation subset;
- In-the-Wild: older real-world deepfake dataset used for evaluation/fine-tuning
Some description of the approach:
The paper evaluates state-of-the-art audio deepfake detectors on real fraud data, studies the benchmark-to-real-world generalization gap, tests threat-informed augmentations and fine-tuning, and compares aggregation strategies for segment-level predictions.
Some description of the data:
The incident data come from online fraud campaigns and fraudulent advertisements. The corpus includes 9,146 URLs, including 7,469 image-based and 1,677 video-based materials, from which the authors extracted over 600 unique audiovisual samples containing manipulated speech.
Keywords:
audio deepfake detection; real-world fraud; incident-based dataset; ASVspoof; In-the-Wild; domain generalization; data augmentation; aggregation strategies.
Instance Represent:
Each instance is represented as either a full audio recording or a segmented 5-second audio clip. For model input, recordings are converted to 16 kHz mono, segmented into non-overlapping 5-second clips, and then trimmed to remove leading and trailing silence longer than 0.2 seconds below 1% peak amplitude. Internal silence is preserved to maintain natural prosody and temporal cues.
Dataset Characteristics:
The study evaluates three real-world datasets: IBD, IBD-partial, and In-the-Wild (ItW). IBD contains 352 verified fraudulent videos and 590 real clips, with 12.5 hours of synthetic audio and 19.3 hours of real audio representing 327 individuals, while IBD-partial contains 34 partial deepfake recordings with authentic and fake segments spliced together. All recordings are standardized to 16 kHz mono, segmented into non-overlapping 5-second clips, and trimmed for long leading/trailing silence while preserving internal silence for prosody and temporal cues.
Subject Area:
Audio deepfake detection, cybersecurity, digital forensics, real-world fraud detection.
Associated Tools:
- Detection models: Mesonet, LCNN, SpecRNet, RawNet3, AASIST.
- Feature/front-end tools: MFCC, LFCC, Wav2Vec 2.0 / W2V2, raw waveform front-end.
- Augmentation tool: audiomentations library .
Feature Type:
The paper analyzes three front-end feature types: MFCC/LFCC classical signal-processing features, Wav2Vec 2.0 self-supervised features, and raw waveform input for end-to-end RawNet3.
Number of Instances:
For the main IBD dataset: 352 fake/fraudulent videos and 590 real clips. After 5-second segmentation and speaker-disjoint split, the test set contains 3,254 samples and the train set contains 16,043 samples.
Number of Features:
Not explicitly stated. The paper uses front-end representations rather than a fixed feature vector: MFCC and LFCC (classical cepstral coefficients), Wav2Vec 2.0 embeddings (self-supervised), and raw waveform input. Specific dimensionalities are not reported.