Konstantin Böttinger’s research while affiliated with Technical University of Munich and other places

... Data allows training deepfake detection methods, as well as measuring their progress. As such, many datasets are continuously proposed [1][2][3][4][5][6][7]. These datasets follow the emergence of recent TTS systems [8][9][10], or aim to extend the number of speakers or languages [5,6]. ...

Reference:

Unmasking real-world audio deepfakes: A data-centric approach

... One alternative approach suggested by Mueller et al. [22] uses Variational Autoencoders to detect latent space 1 github.com/Arsu-Lab/Shortcut-Detection-Mitigation-Transformers dimensions with high label predictiveness and generate a set of images that differ in a single image attribute. ...

Reference:

Efficient Unsupervised Shortcut Learning Detection and Mitigation in Transformers

... Bhagtani et al. [11] compared several recently proposed detection tools on their new dataset and demonstrated that synthetic speech detectors show significant drops in accuracy compared to their performance on training datasets. Similarly, Muller et al. [57] evaluated the performance of various detection tools on inand out-of-domain test data. Their findings confirm that these existing detection tools do not generalize well to unseen data. ...

Reference:

VoiceRadar: Voice Deepfake Detection using Micro-Frequency and Compositional Analysis

... The objective of this section is to analyse some of the threats posed by deepfakes and outline their mitigation methods. Identifying deepfakes is the first step, using deep learning to detect digital artifacts like facial inconsistencies (Müller, N. et al., 2023) The next step is to protect video surveillance systems from hacking, employing biometric authentication, encryption, and network monitoring to secure data (Vennam, P. et al., 2021). The final part of this chapter discusses algorithms for cyberattack detection, such as CNNs and RNNs, which help detect anomalies in video frames and patterns of suspicious activity (Wodajo, D. & Atnafu, S., 2021). ...

Reference:

Application of Computer Vision Methods for Information Security

... The observed performance of the models on the data sets is largely consistent with results from earlier studies conducted by other authors [37], [38]. Differences in performance between the data sets are likely attributable to the varying complexities present in each data set. ...

Reference:

Semisupervised Anomaly Detection using Support Vector Regression with Quantum Kernel

... Adversarial weather attacks. In contrast, adversarial attacks that imitate weather effects have been investigated for classification [7,8,16,23,53], object detection [8,35,52], instance segmentation [8], human pose estimation [44] or autonomous steering [21]. They range from rain [8,23,35,52] over snow [8,16,23] to fog [7,16,21] and shadows [53]. ...

Reference:

Distracting Downpour: Adversarial Weather Attacks for Motion Estimation

... Pizzi et al. [83] demonstrated the feasibility of model inversion attacks on speaker recognition by directly attacking the target model to regenerate representative audio samples and extract sensitive speaker embeddings learned by the model. ...

Reference:

Adversarial Attacks in Multimodal Systems: A Practitioner's Survey

... For example, the entropy of a given layer, as well as the closely-related mutual information, have recently been evaluated via replica methods in [32]. Entropy has also been proposed as a training mechanism in the context of adversarial learning in [33]; see also [34,35]. Additionally, estimates of f -divergences (of which the KL divergence is perhaps the canonical example) are relevant for generative adversarial networks, see e.g., [36]. ...

Reference:

Towards quantifying information flows: relative entropy in deep neural networks and the renormalization group

... The effectiveness of GAN-based techniques, which use a generative adversarial network, can be reduced without any knowledge of the attack (Samangouei et al., 2018). From gradient masking techniques, which modify the gradients of input data, both gradient penalization (Lyu et al., 2016) and loss function minimization (Muller et al., 2021) cannot defend well against black-box attacks. Gradient regularization techniques, which add constraints to the objective function during training, cannot improve significantly the robustness of the neural network without reducing the accuracy of the model (Moosavi-Dezfooli, 2019). ...

Reference:

European Artificial Intelligence Act: an AI security approach

... Our experiments utilize multiple diverse datasets containing bonafide and synthetically spoofed audio samples that were created using various synthesis methods. We specifically selected datasets that reflect distinct class distributions, thereby allowing comprehensive evaluation under realistic scenarios: ASVspoof2019 [20] is significantly skewed toward spoof samples, making it challenging for models to correctly identify genuine audio; In-the-Wild [21] predominantly comprises bonafide samples, testing the models' sensitivity and specificity in a practical context; and WaveFake [22] presents a balanced dataset equally composed of spoofed and bonafide samples, providing a controlled setting for evaluating model generalization and robustness. Table 1 summarizes key dataset characteristics. ...

Reference:

Can Quantized Audio Language Models Perform Zero-Shot Spoofing Detection?