The emergence of federated learning (FL) promotes the rapid development of artificial intelligence technology in Internet of Things (IoT). However, the huge communication cost and consumption of local computing resource in the process of FL restricts its further development and application in IoT. To solve the above problems, this paper proposes an efficient and secure self-learning federated learning method based on chaotic secure clustering. Firstly, the self-learning weights selection strategy is proposed, this strategy divides FL into two stages. In the first stage, the clients randomly select partial weights according to the inactivation parameter, which are encrypted by CKKS homomorphic encryption and uploaded, and the server updates the weights adaptive matrix according to the performance of the global model. In the second stage, the clients select some weights to encrypt and upload according to the weights adaptive matrix and inactivation parameter. Then, to further improve the efficiency of FL, we propose the chaotic secure clustering method to select some clients. The server performs clustering operation before FL. And in the clustering process, we transform and locally-globally scramble the client label information, which ensures the security of the client information in the clustering process. Finally, we use MNIST dataset, fashion-MNIST dataset, GTSRB dataset and CSE-CIC-IDS2018 dataset to simulate non-independent and identically distributed (non-IID) scenarios, and the experimental results show the effectiveness, security and efficiency of the proposed method, and our method provides a new learning paradigm for the further development and application of FL in IoT.

With the development of content generative large models, spoofed speech can be synthesized more easily. However, the generalization of current cloud voice anti-spoofing detection model is insufficient, especially when facing attacks from unknown speech synthesis algorithms. Meanwhile, voice data is also susceptible to attacks during transmission. Therefore, we design an end-to-end voice anti-spoofing scheme, which can be applied in IoT cloud service. The scheme consists of an encryption transmission module and an anti-spoofing model. The encryption module is designed with post-quantum cryptography and chaos to protect transmission security. In the modle, we propose a new higher-order two-dimensional attentive statistics pooling (H2D-ASP) module to extract and aggregate more attention representations in spectral domain and temporal domain; And we propose a new channel-dependent self attention based graph aggregation (CSA-GA) module, which squeezes and aggregates spectral graphs and temporal graphs. Finally, we conduct experiments on the ASVspoof 5 Challenge deepfake database under the closed condition. The experiments show that the model in our scheme is a better single model which performs minDCF 18.28% better than the baseline model on the evaluation set. Without data augmentation, the model achieves minDCF of 0.2994, 0.3604, 0.581 on the development set, evaluationprog set and evaluation set, respectively. The last two proposed modules improve the vanilla model by 29.75% on evaluationprog set and 18.97% on evaluation set.

Voice spoofing detection is a technique for enhancing the security of automatic speaker verification system, but the existing research still faces problems such as weak detection capability and expensive computation. To address these problems, this work presents a lightweight voice anti-spoofing method by using improved one-class learning DOC-Softmax and knowledge distillation. The main idea of DOC-Softmax is to learn a feature space where the genuine samples have a compact space and the spoofing samples are parted from the bona fide space by a certain interval. And the dispersion loss is introduced for spoofing samples to cover the whole spoofing space as much as possible. Moreover, a lightweight voice spoofing detection model is designed to speed up inference, and the knowledge distillation is employed to improve representation power of the lightweight model. Without any data augmentation and ensemble learning, a series of experiments are conducted on LA and PA scenarios of the ASVspoof 2019 dataset, and the experimental results indicate that the proposed method performs better than most existing voice anti-spoofing methods.

Most of the voice spoofing detection methods are designed for specific kinds of spoofing attacks, synthetic or replay. In practice, however, there is no prior information about these two kinds of spoofing attacks. To this end, this paper proposes a generalized voice spoofing detection method based on integral knowledge amalgamation to detect jointly synthetic attacks and replay attacks. Two amalgamation mechanisms, feature amalgamation and structure amalgamation, are designed from different perspectives, so that the model can generalize better and run fast. Specifically, the feature amalgamation transfers the high-level sematic knowledge from two teacher models to the compact model. The structure amalgamation employs the adversarial learning to ensure the global structure consistency of two teacher models and a student model. In addition, the feature matching loss is introduced to capture the distinctive features of synthetic attacks and replay attacks. We conduct extensive experiments on logical access (LA) scenario and physical access (PA) scenario of ASVspoof 2019 dataset to verify the validity of the proposed method. The experimental results show that compared with the most advanced generalized voice spoofing detection methods, the proposed method achieves a comparable or even better performance. In particular, our method gains the state-of-the-art detection capability on LA scenario. Moreover, our method achieves similar or even outstanding detectability when compared with specialized anti-spoofing methods.