Xiang Li

Diffusion models have gained prominence in the image domain for their capabilities in data generation and transformation, achieving state-of-the-art performance in various tasks in both image and audio domains. In the rapidly evolving field of audio-based machine learning, safeguarding model integrity and establishing data copyright are of paramount importance. This paper presents the first watermarking technique applied to audio diffusion models trained on mel-spectrograms. This offers a novel approach to the aforementioned challenges. Our model excels not only in benign audio generation, but also incorporates an invisible watermarking trigger mechanism for model verification. This watermark trigger serves as a protective layer, enabling the identification of model ownership and ensuring its integrity. Through extensive experiments, we demonstrate that invisible watermark triggers can effectively protect against unauthorized modifications while maintaining high utility in benign audio generation tasks.

In clinical practice, one patient may suffer from more than one arrhythmia simultaneously, that is, one ECG record may be associated with multiple types of arrhythmias. In fact, there are inherent dependencies between arrhythmias. However, previous studies have mainly focused on multi-class (single-label) ECG classification, which addresses each type of arrhythmia independently and ignores the multi-label correlation between different ECG abnormalities. To address the lack of ECG multi-label classification methods, we proposed a label correlation embedding guided network (LCEGNet) model to effectively recognize multi-label ECG arrhythmias and explore the correlation between ECG abnormalities. First, label correlation embedding was obtained based on the correlation matrix between different arrhythmias to guide feature extraction. Subsequently, the category-specific attention coefficient was obtained by calculating the cosine similarity coefficient between the label embedding and feature spaces. Experiments on public and self-collected ECG datasets were conducted. The LCEGNet achieved F1 scores of 0.777 and 0.872 and subset accuracy of 0.750 and 0.828 on the two datasets, respectively. A classification speed of 7.796 ms was achieved. The experimental results demonstrate that the proposed LCEGNet achieved approximately a 11% and 9.1% improvement in the F1 score and subset accuracy, respectively, compared with traditional ResNet architecture and a 4.3% and 5.54% improvement in the F1 score and subset accuracy, respectively, compared with the state-of-the-art approaches.