An official website of the United States government
Cochlear implants (CIs) allow deaf and hard-ofhearing individuals to use audio devices, such as phones or voice assistants. However, the advent of increasingly sophisticated synthetic audio (i.e., deepfakes) potentially threatens these users. Yet, this population’s susceptibility to such attacks is unclear. In this paper, we perform the first study of the impact of audio deepfakes on CI populations. We examine the use of CI-simulated audio within deepfake detectors. Based on these results, we conduct a user study with 35 CI users and 87 hearing persons (HPs) to determine differences in how CI users perceive deepfake audio. We show that CI users can, similarly to HPs, identify text-to-speech generated deepfakes. Yet, they perform substantially worse for voice conversion deepfake generation algorithms, achieving only 67% correct audio classification. We also evaluate how detection models trained on a CI-simulated audio compare to CI users and investigate if they can effectively act as proxies for CI users. This work begins an investigation into the intersection between adversarial audio and CI users to identify and mitigate threats against this marginalized group.
more »
« less
This content will become publicly available on September 1, 2026
Every Breath You Don’t Take: Deepfake Speech Detection Using Breath
Deepfake speech represents a real and growing threat to systems and society. Many detectors have been created to aid in defense against speech deepfakes. While these detectors implement myriad methodologies, many rely on low-level fragments of the speech generation process. We hypothesize that breath, a higher-level part of speech, is a key component of natural speech and thus improper generation in deepfake speech is a performant discriminator. To evaluate this, we create a breath detector and leverage this against a custom dataset of online news article audio to discriminate between real/deepfake speech. Additionally, we make this custom dataset publicly available to facilitate comparison for future work. Applying our simple breath detector as a deepfake speech discriminator on in-the-wild samples allows for accurate classification (perfect 1.0 AUPRC and 0.0 EER on test data) across 33.6 hours of audio. We compare our model with the state-of-the-art SSL-wav2vec and Codecfake models and show that these complex deep learning model completely either fail to classify the same in-the-wild samples (0.72 AUPRC and 0.89 EER), or substantially lack in the computational and temporal performance compared to our methodology (37 seconds to predict a one minute sample with Codecfake vs. 0.3 seconds with our model)
more »
« less
- Award ID(s):
- 1933208
- PAR ID:
- 10616855
- Publisher / Repository:
- ACM
- Date Published:
- Journal Name:
- ACM Digital Threats: Research and Practice (DTRAP)
- ISSN:
- 2576-5338
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
We present an innovative approach to auto-annotate Expert Defined Linguistic Features (EDLFs) as subsequences in audio time series to improve audio deepfake discernment. In our prior work, these linguistic features – namely pitch, pause, breath, consonant release bursts, and overall audio quality, labeled by experts on the entire audio signal – have been shown to improve detection of audio deepfakes with AI algorithms. We now expand our approach to pilot a way to auto annotate subsequences in the time series that correspond to each EDLF. We developed an ensemble of discords, i.e. anomalies in time series, detected using matrix profiles across multiple discord lengths to identify multiple types of EDLFs. Working closely with linguistic experts, we evaluated where discords overlapped with EDLFs in the audio signal data. Our ensemble method to detect discords across multiple discord lengths achieves much higher accuracy than using individual discord lengths to detect EDLFs. With this approach and domain validation we establish the feasibility of using time series subsequences to capture EDLFs to supplement annotation by domain experts, for improved audio deepfake detection.more » « less
-
The rapid development of deep neural networks and generative AI has catalyzed growth in realistic speech synthesis. While this technology has great potential to improve lives, it also leads to the emergence of ''DeepFake'' where synthesized speech can be misused to deceive humans and machines for nefarious purposes. In response to this evolving threat, there has been a significant amount of interest in mitigating this threat by DeepFake detection. Complementary to the existing work, we propose to take the preventative approach and introduce AntiFake, a defense mechanism that relies on adversarial examples to prevent unauthorized speech synthesis. To ensure the transferability to attackers' unknown synthesis models, an ensemble learning approach is adopted to improve the generalizability of the optimization process. To validate the efficacy of the proposed system, we evaluated AntiFake against five state-of-the-art synthesizers using real-world DeepFake speech samples. The experiments indicated that AntiFake achieved over 95% protection rate even to unknown black-box models. We have also conducted usability tests involving 24 human participants to ensure the solution is accessible to diverse populations.more » « less
-
Collaboration is a 21st Century skill as well as an effective method for learning, so detection of collaboration is important for both assessment and instruction. Speech-based collaboration detection can be quite accurate but collecting the speech of students in classrooms can raise privacy issues. An alternative is to send only whether or not the student is speaking. That is, the speech signal is processed at the microphone by a voice activity detector before being transmitted to the collaboration detector. Because the transmitted signal is binary (1 = speaking, 0 = silence), this method mitigates privacy issues. However, it may harm the accuracy of collaboration detection. To find out how much harm is done, this study compared the relative effectiveness of collaboration detectors based either on the binary signal or high-quality audio. Pairs of students were asked to work together on solving complex math problems. Three qualitative levels of interactivity was distinguished: Interaction, Cooperation and Other. Human coders used richer data (several audio and video streams) to choose the code for each episode. Machine learning was used to induce a detector to assign a code for every episode based on the features. The binary-based collaboration detectors delivered only slightly less accuracy than collaboration detectors based on the high quality audio signal.more » « less
-
The robustness and vulnerability of Deep Neural Networks (DNN) are quickly becoming a critical area of interest since these models are in widespread use across real-world applications (i.e., image and audio analysis, recommendation system, natural language analysis, etc.). A DNN's vulnerability is exploited by an adversary to generate data to attack the model; however, the majority of adversarial data generators have focused on image domains with far fewer work on audio domains. More recently, audio analysis models were shown to be vulnerable to adversarial audio examples (e.g., speech command classification, automatic speech recognition, etc.). Thus, one urgent open problem is to detect adversarial audio reliably. In this contribution, we incorporate a separate and yet related DNN technique to detect adversarial audio, namely model quantization. Then we propose an algorithm to detect adversarial audio by using a DNN's quantization error. Specifically, we demonstrate that adversarial audio typically exhibits a larger activation quantization error than benign audio. The quantization error is measured using character error rates. We use the difference in errors to discriminate adversarial audio. Experiments with three the-state-of-the-art audio attack algorithms against the DeepSpeech model show our detection algorithm achieved high accuracy on the Mozilla dataset.more » « less
