Nicolas M. Müller

Nicolas M. Müller, Piotr Kawa, Wei Herng Choong et al.

This paper presents the Multi-Language Audio Anti-Spoofing Dataset (MLAAD), version 10: a dataset of synthetic audio to train and evaluate audio deepfake detection models. It features 175 Text-to-Speech (TTS) models, comprising a total of 1002.9 hours of synthetic voice in 54 different languages. To evaluate this dataset, we train three state-of-the-art deepfake detection models with MLAAD and observe that it demonstrates superior performance to comparable datasets like InTheWild and FakeOrReal when used as a training resource. Moreover, compared to the renowned ASVspoof 2019 dataset, MLAAD proves to be a complementary resource. In tests across eight datasets, MLAAD and ASVspoof 2019 alternately outperformed each other, each excelling on four datasets. By publishing the dataset and making a trained model accessible via an interactive webserver, we aim to democratize anti-spoofing technology, making it accessible beyond the realm of specialists, and contributing to global efforts against audio spoofing and deepfakes.

Octavian Pascu, Dan Oneata, Horia Cucu et al.

In this paper, we demonstrate that attacks in the latest ASVspoof5 dataset -- a de facto standard in the field of voice authenticity and deepfake detection -- can be identified with surprising accuracy using a small subset of very simplistic features. These are derived from the openSMILE library, and are scalar-valued, easy to compute, and human interpretable. For example, attack A10`s unvoiced segments have a mean length of 0.09 +- 0.02, while bona fide instances have a mean length of 0.18 +- 0.07. Using this feature alone, a threshold classifier achieves an Equal Error Rate (EER) of 10.3% for attack A10. Similarly, across all attacks, we achieve up to 0.8% EER, with an overall EER of 15.7 +- 6.0%. We explore the generalization capabilities of these features and find that some of them transfer effectively between attacks, primarily when the attacks originate from similar Text-to-Speech (TTS) architectures. This finding may indicate that voice anti-spoofing is, in part, a problem of identifying and remembering signatures or fingerprints of individual TTS systems. This allows to better understand anti-spoofing models and their challenges in real-world application.

Nicolas M. Müller, Philip Sperl, Konstantin Böttinger

Current anti-spoofing and audio deepfake detection systems use either magnitude spectrogram-based features (such as CQT or Melspectrograms) or raw audio processed through convolution or sinc-layers. Both methods have drawbacks: magnitude spectrograms discard phase information, which affects audio naturalness, and raw-feature-based models cannot use traditional explainable AI methods. This paper proposes a new approach that combines the benefits of both methods by using complex-valued neural networks to process the complex-valued, CQT frequency-domain representation of the input audio. This method retains phase information and allows for explainable AI methods. Results show that this approach outperforms previous methods on the "In-the-Wild" anti-spoofing dataset and enables interpretation of the results through explainable AI. Ablation studies confirm that the model has learned to use phase information to detect voice spoofing.

Nicolas M. Müller, Franziska Dieckmann, Jennifer Williams

Deepfakes are synthetically generated media often devised with malicious intent. They have become increasingly more convincing with large training datasets advanced neural networks. These fakes are readily being misused for slander, misinformation and fraud. For this reason, intensive research for developing countermeasures is also expanding. However, recent work is almost exclusively limited to deepfake detection - predicting if audio is real or fake. This is despite the fact that attribution (who created which fake?) is an essential building block of a larger defense strategy, as practiced in the field of cybersecurity for a long time. This paper considers the problem of deepfake attacker attribution in the domain of audio. We present several methods for creating attacker signatures using low-level acoustic descriptors and machine learning embeddings. We show that speech signal features are inadequate for characterizing attacker signatures. However, we also demonstrate that embeddings from a recurrent neural network can successfully characterize attacks from both known and unknown attackers. Our attack signature embeddings result in distinct clusters, both for seen and unseen audio deepfakes. We show that these embeddings can be used in downstream-tasks to high-effect, scoring 97.10% accuracy in attacker-id classification.

Nicolas M. Müller, Jochen Jacobs, Jennifer Williams et al.

Machine learning is a data-driven field, and the quality of the underlying datasets plays a crucial role in learning success. However, high performance on held-out test data does not necessarily indicate that a model generalizes or learns anything meaningful. This is often due to the existence of machine learning shortcuts - features in the data that are predictive but unrelated to the problem at hand. To address this issue for datasets where the shortcuts are smaller and more localized than true features, we propose a novel approach to detect and remove them. We use an adversarially trained lens to detect and eliminate highly predictive but semantically unconnected clues in images. In our experiments on both synthetic and real-world data, we show that our proposed approach reliably identifies and neutralizes such shortcuts without causing degradation of model performance on clean data. We believe that our approach can lead to more meaningful and generalizable machine learning models, especially in scenarios where the quality of the underlying datasets is crucial.

Octavian Pascu, Dan Oneata, Horia Cucu et al.

We introduce Echoes, a new dataset for music deepfake detection designed for training and benchmarking detectors under realistic and provider-diverse conditions. Echoes comprises 3,577 tracks (110 hours of audio) spanning multiple genres (pop, rock, electronic), and includes content generated by ten popular AI music generation systems. To prevent shortcut learning and promote robust generalization, the dataset is deliberately constructed to be challenging, enforcing semantic-level alignment between spoofed audio and bona fide references. This alignment is achieved by conditioning generated audio samples directly on bona-fide waveforms or song descriptors. We evaluate Echoes in a cross-dataset setting against three existing AI-generated music datasets using state-of-the-art Wav2Vec2 XLS-R 2B representations. Results show that (i) Echoes is the hardest in-domain dataset; (ii) detectors trained on existing datasets transfer poorly to Echoes; (iii) training on Echoes yields the strongest generalization performance. These findings suggest that provider diversity and semantic alignment help learn more transferable detection cues.

Nicolas M. Müller, Simon Roschmann, Shahbaz Khan et al.

For real-world applications of machine learning (ML), it is essential that models make predictions based on well-generalizing features rather than spurious correlations in the data. The identification of such spurious correlations, also known as shortcuts, is a challenging problem and has so far been scarcely addressed. In this work, we present a novel approach to detect shortcuts in image and audio datasets by leveraging variational autoencoders (VAEs). The disentanglement of features in the latent space of VAEs allows us to discover feature-target correlations in datasets and semi-automatically evaluate them for ML shortcuts. We demonstrate the applicability of our method on several real-world datasets and identify shortcuts that have not been discovered before.

Nicolas M. Müller, Maximilian Burgert, Pascal Debus et al.

Machine-learning (ML) shortcuts or spurious correlations are artifacts in datasets that lead to very good training and test performance but severely limit the model's generalization capability. Such shortcuts are insidious because they go unnoticed due to good in-domain test performance. In this paper, we explore the influence of different shortcuts and show that even simple shortcuts are difficult to detect by explainable AI methods. We then exploit this fact and design an approach to defend online databases against crawlers: providers such as dating platforms, clothing manufacturers, or used car dealers have to deal with a professionalized crawling industry that grabs and resells data points on a large scale. We show that a deterrent can be created by deliberately adding ML shortcuts. Such augmented datasets are then unusable for ML use cases, which deters crawlers and the unauthorized use of data from the internet. Using real-world data from three use cases, we show that the proposed approach renders such collected data unusable, while the shortcut is at the same time difficult to notice in human perception. Thus, our proposed approach can serve as a proactive protection against illegitimate data crawling.

Nicolas Müller, Piotr Kawa, Wei-Herng Choong et al.

We show how replay attacks undermine audio deepfake detection: By playing and re-recording deepfake audio through various speakers and microphones, we make spoofed samples appear authentic to the detection model. To study this phenomenon in more detail, we introduce ReplayDF, a dataset of recordings derived from M-AILABS and MLAAD, featuring 109 speaker-microphone combinations across six languages and four TTS models. It includes diverse acoustic conditions, some highly challenging for detection. Our analysis of six open-source detection models across five datasets reveals significant vulnerability, with the top-performing W2V2-AASIST model's Equal Error Rate (EER) surging from 4.7% to 18.2%. Even with adaptive Room Impulse Response (RIR) retraining, performance remains compromised with an 11.0% EER. We release ReplayDF for non-commercial research use.

Nicolas M. Müller, Piotr Kawa, Shen Hu et al.

Voice faking, driven primarily by recent advances in text-to-speech (TTS) synthesis technology, poses significant societal challenges. Currently, the prevailing assumption is that unaltered human speech can be considered genuine, while fake speech comes from TTS synthesis. We argue that this binary distinction is oversimplified. For instance, altered playback speeds can be used for malicious purposes, like in the 'Drunken Nancy Pelosi' incident. Similarly, editing of audio clips can be done ethically, e.g., for brevity or summarization in news reporting or podcasts, but editing can also create misleading narratives. In this paper, we propose a conceptual shift away from the binary paradigm of audio being either 'fake' or 'real'. Instead, our focus is on pinpointing 'voice edits', which encompass traditional modifications like filters and cuts, as well as TTS synthesis and VC systems. We delineate 6 categories and curate a new challenge dataset rooted in the M-AILABS corpus, for which we present baseline detection systems. And most importantly, we argue that merely categorizing audio as fake or real is a dangerous over-simplification that will fail to move the field of speech technology forward.

Daniel Kowatsch, Nicolas M. Müller, Kilian Tscharke et al.

For classification, the problem of class imbalance is well known and has been extensively studied. In this paper, we argue that imbalance in regression is an equally important problem which has so far been overlooked: Due to under- and over-representations in a data set's target distribution, regressors are prone to degenerate to naive models, systematically neglecting uncommon training data and over-representing targets seen often during training. We analyse this problem theoretically and use resulting insights to develop a first definition of imbalance in regression, which we show to be a generalisation of the commonly employed imbalance measure in classification. With this, we hope to turn the spotlight on the overlooked problem of imbalance in regression and to provide common ground for future research.

Nicolas Müller, Piotr Kawa, Adriana Stan et al.

Deepfakes - manipulated or forged audio and video media - pose significant security risks to individuals, organizations, and society at large. To address these challenges, machine learning-based classifiers are commonly employed to detect deepfake content. In this paper, we assess the robustness of such classifiers through a systematic penetration testing methodology, which we introduce as DeePen. Our approach operates without prior knowledge of or access to the target deepfake detection models. Instead, it leverages a set of carefully selected signal processing modifications - referred to as attacks - to evaluate model vulnerabilities. Using DeePen, we analyze both real-world production systems and publicly available academic model checkpoints, demonstrating that all tested systems exhibit weaknesses and can be reliably deceived by simple manipulations such as time-stretching or echo addition. Furthermore, our findings reveal that while some attacks can be mitigated by retraining detection systems with knowledge of the specific attack, others remain persistently effective. We release all associated code.

Nicolas M. Müller, Nicholas Evans, Hemlata Tak et al.

Recent research has highlighted a key issue in speech deepfake detection: models trained on one set of deepfakes perform poorly on others. The question arises: is this due to the continuously improving quality of Text-to-Speech (TTS) models, i.e., are newer DeepFakes just 'harder' to detect? Or, is it because deepfakes generated with one model are fundamentally different to those generated using another model? We answer this question by decomposing the performance gap between in-domain and out-of-domain test data into 'hardness' and 'difference' components. Experiments performed using ASVspoof databases indicate that the hardness component is practically negligible, with the performance gap being attributed primarily to the difference component. This has direct implications for real-world deepfake detection, highlighting that merely increasing model capacity, the currently-dominant research trend, may not effectively address the generalization challenge.

Nicolas M. Müller, Pavel Czempin, Franziska Dieckmann et al.

Current text-to-speech algorithms produce realistic fakes of human voices, making deepfake detection a much-needed area of research. While researchers have presented various techniques for detecting audio spoofs, it is often unclear exactly why these architectures are successful: Preprocessing steps, hyperparameter settings, and the degree of fine-tuning are not consistent across related work. Which factors contribute to success, and which are accidental? In this work, we address this problem: We systematize audio spoofing detection by re-implementing and uniformly evaluating architectures from related work. We identify overarching features for successful audio deepfake detection, such as using cqtspec or logspec features instead of melspec features, which improves performance by 37% EER on average, all other factors constant. Additionally, we evaluate generalization capabilities: We collect and publish a new dataset consisting of 37.9 hours of found audio recordings of celebrities and politicians, of which 17.2 hours are deepfakes. We find that related work performs poorly on such real-world data (performance degradation of up to one thousand percent). This may suggest that the community has tailored its solutions too closely to the prevailing ASVSpoof benchmark and that deepfakes are much harder to detect outside the lab than previously thought.

Nicolas M. Müller, Karla Pizzi, Jennifer Williams

The recent emergence of deepfakes has brought manipulated and generated content to the forefront of machine learning research. Automatic detection of deepfakes has seen many new machine learning techniques, however, human detection capabilities are far less explored. In this paper, we present results from comparing the abilities of humans and machines for detecting audio deepfakes used to imitate someone's voice. For this, we use a web-based application framework formulated as a game. Participants were asked to distinguish between real and fake audio samples. In our experiment, 472 unique users competed against a state-of-the-art AI deepfake detection algorithm for 14912 total of rounds of the game. We find that humans and deepfake detection algorithms share similar strengths and weaknesses, both struggling to detect certain types of attacks. This is in contrast to the superhuman performance of AI in many application areas such as object detection or face recognition. Concerning human success factors, we find that IT professionals have no advantage over non-professionals but native speakers have an advantage over non-native speakers. Additionally, we find that older participants tend to be more susceptible than younger ones. These insights may be helpful when designing future cybersecurity training for humans as well as developing better detection algorithms.

Nicolas M. Müller, Franziska Dieckmann, Pavel Czempin et al.

We present our analysis of a significant data artifact in the official 2019/2021 ASVspoof Challenge Dataset. We identify an uneven distribution of silence duration in the training and test splits, which tends to correlate with the target prediction label. Bonafide instances tend to have significantly longer leading and trailing silences than spoofed instances. In this paper, we explore this phenomenon and its impact in depth. We compare several types of models trained on a) only the duration of the leading silence and b) only on the duration of leading and trailing silence. Results show that models trained on only the duration of the leading silence perform particularly well, and achieve up to 85% percent accuracy and an equal error rate (EER) of 15.1%. At the same time, we observe that trimming silence during pre-processing and then training established antispoofing models using signal-based features leads to comparatively worse performance. In that case, EER increases from 3.6% (with silence) to 15.5% (trimmed silence). Our findings suggest that previous work may, in part, have inadvertently learned thespoof/bonafide distinction by relying on the duration of silence as it appears in the official challenge dataset. We discuss the potential consequences that this has for interpreting system scores in the challenge and discuss how the ASV community may further consider this issue.

Nicolas M. Müller, Simon Roschmann, Konstantin Böttinger

Data poisoning is one of the most relevant security threats against machine learning and data-driven technologies. Since many applications rely on untrusted training data, an attacker can easily craft malicious samples and inject them into the training dataset to degrade the performance of machine learning models. As recent work has shown, such Denial-of-Service (DoS) data poisoning attacks are highly effective. To mitigate this threat, we propose a new approach of detecting DoS poisoned instances. In comparison to related work, we deviate from clustering and anomaly detection based approaches, which often suffer from the curse of dimensionality and arbitrary anomaly threshold selection. Rather, our defence is based on extracting information from the training data in such a generalized manner that we can identify poisoned samples based on the information present in the unpoisoned portion of the data. We evaluate our defence against two DoS poisoning attacks and seven datasets, and find that it reliably identifies poisoned instances. In comparison to related work, our defence improves false positive / false negative rates by at least 50%, often more.

Edresson Casanova, Christopher Shulby, Eren Gölge et al.

In this paper, we propose SC-GlowTTS: an efficient zero-shot multi-speaker text-to-speech model that improves similarity for speakers unseen during training. We propose a speaker-conditional architecture that explores a flow-based decoder that works in a zero-shot scenario. As text encoders, we explore a dilated residual convolutional-based encoder, gated convolutional-based encoder, and transformer-based encoder. Additionally, we have shown that adjusting a GAN-based vocoder for the spectrograms predicted by the TTS model on the training dataset can significantly improve the similarity and speech quality for new speakers. Our model converges using only 11 speakers, reaching state-of-the-art results for similarity with new speakers, as well as high speech quality.

Pascal Debus, Nicolas Müller, Konstantin Böttinger

Many defensive measures in cyber security are still dominated by heuristics, catalogs of standard procedures, and best practices. Considering the case of data backup strategies, we aim towards mathematically modeling the underlying threat models and decision problems. By formulating backup strategies in the language of stochastic processes, we can translate the challenge of finding optimal defenses into a reinforcement learning problem. This enables us to train autonomous agents that learn to optimally support planning of defense processes. In particular, we tackle the problem of finding an optimal backup scheme in the following adversarial setting: Given $k$ backup devices, the goal is to defend against an attacker who can infect data at one time but chooses to destroy or encrypt it at a later time, potentially also corrupting multiple backups made in between. In this setting, the usual round-robin scheme, which always replaces the oldest backup, is no longer optimal with respect to avoidable exposure. Thus, to find a defense strategy, we model the problem as a hybrid discrete-continuous action space Markov decision process and subsequently solve it using deep deterministic policy gradients. We show that the proposed algorithm can find storage device update schemes which match or exceed existing schemes with respect to various exposure metrics.

Nicolas M. Müller, Konstantin Böttinger

Two widely used techniques for training supervised machine learning models on small datasets are Active Learning and Transfer Learning. The former helps to optimally use a limited budget to label new data. The latter uses large pre-trained models as feature extractors and enables the design of complex, non-linear models even on tiny datasets. Combining these two approaches is an effective, state-of-the-art method when dealing with small datasets. In this paper, we share an intriguing observation: Namely, that the combination of these techniques is particularly susceptible to a new kind of data poisoning attack: By adding small adversarial noise on the input, it is possible to create a collision in the output space of the transfer learner. As a result, Active Learning algorithms no longer select the optimal instances, but almost exclusively the ones injected by the attacker. This allows an attacker to manipulate the active learner to select and include arbitrary images into the data set, even against an overwhelming majority of unpoisoned samples. We show that a model trained on such a poisoned dataset has a significantly deteriorated performance, dropping from 86\% to 34\% test accuracy. We evaluate this attack on both audio and image datasets and support our findings empirically. To the best of our knowledge, this weakness has not been described before in literature.

Tom Dörr, Karla Markert, Nicolas M. Müller et al.

Adversarial examples tremendously threaten the availability and integrity of machine learning-based systems. While the feasibility of such attacks has been observed first in the domain of image processing, recent research shows that speech recognition is also susceptible to adversarial attacks. However, reliably bridging the air gap (i.e., making the adversarial examples work when recorded via a microphone) has so far eluded researchers. We find that due to flaws in the generation process, state-of-the-art adversarial example generation methods cause overfitting because of the binning operation in the target speech recognition system (e.g., Mozilla Deepspeech). We devise an approach to mitigate this flaw and find that our method improves generation of adversarial examples with varying offsets. We confirm the significant improvement with our approach by empirical comparison of the edit distance in a realistic over-the-air setting. Our approach states a significant step towards over-the-air attacks. We publish the code and an applicable implementation of our approach.

Nicolas Michael Müller, Daniel Kowatsch, Konstantin Böttinger

Adversarial data poisoning is an effective attack against machine learning and threatens model integrity by introducing poisoned data into the training dataset. So far, it has been studied mostly for classification, even though regression learning is used in many mission critical systems (such as dosage of medication, control of cyber-physical systems and managing power supply). Therefore, in the present research, we aim to evaluate all aspects of data poisoning attacks on regression learning, exceeding previous work both in terms of breadth and depth. We present realistic scenarios in which data poisoning attacks threaten production systems and introduce a novel black-box attack, which is then applied to a real-word medical use-case. As a result, we observe that the mean squared error (MSE) of the regressor increases to 150 percent due to inserting only two percent of poison samples. Finally, we present a new defense strategy against the novel and previous attacks and evaluate it thoroughly on 26 datasets. As a result of the conducted experiments, we conclude that the proposed defence strategy effectively mitigates the considered attacks.