Voiceprint verification has been widely adopted in daily applications due to its convenience and efficiency. However, the latent representations extracted during speech verification often encode sensitive personal information, such as emotion and gender, which raises significant privacy concerns. Existing methods mainly focus on protecting predefined sensitive attributes, but in many cases, the attributes that attackers are interested in are unknown to defenders. To address this limitation, we propose the Speech Information Bottleneck Network (SIBNet), a framework for representation level privacy enhancement based on information theoretic approaches. SIBNet differs from existing speech attribute privacy mechanisms in three key respects: (1) We formulate biometric privacy protection as an Information Bottleneck (IB) optimization problem that balances identity utility and sensitive attribute suppression. (2) SIBNet leverages a variational distribution parameterized by neural networks to approximate mutual information, enabling the derivation of lower bounds for effective model optimization. (3) An additive angular margin penalty is introduced into the speaker classification loss to promote discriminative embeddings and preserve identity utility. Experiments on public speech datasets demonstrate that SIBNet significantly outperforms state of the art methods in both biometric recognition accuracy and suppression of sensitive soft biometric attributes.

Abstract In recent years, the proliferation of facial image collection systems coupled with significant advancements in machine learning-driven image analysis techniques has posed formidable challenges to protecting individuals’ privacy information, raising concerns about the security of such sensitive data. The current state-of-the-art technology is adept at extracting an array of intimate personal privacy details, encompassing gender, race, and potentially more, from a solitary facial image, underscoring the intricacies and implications of data privacy. Therefore, there is an urgent need for research on models that can protect the privacy of facial images. To tackle this issue, we proposed a multi-attribute privacy-preserving computational model based on generative adversarial networks (MAP-GAN) to protect sensitive facial privacy attribute information at the image level. For MAP-GAN, we meticulously design a privacy preservation loss function and introduce an attribute probability scoring mechanism to address the problem of binary attribute privacy protection flipping observed in previous models. Additionally, we incorporate an $$\textit{L}_1$$ L 1 information loss constraint to ensure the practical superiority of MAP-GAN by maintaining the information gap between privacy-protected images and original images. To enhance the quality of generated privacy-preserving images, we present a privacy-preserving image generator that utilizes residual structures and selective transmission units in MAP-GAN’s design. Experimental results demonstrate that MAP-GAN outperforms other models in terms of both multi-attribute privacy protection and utility.

Consumer electronic devices with voice assistants are becoming increasingly popular in modern intelligent home. Nevertheless, directly uploading unprocessed speech data, which may contain sensitive attributes, to a cloud server poses a significant risk to user privacy. To address this privacy issue, this paper proposes a privacy-enhancing model to protect speech emotions based on generative adversarial networks (PSEGAN). The model aims to prevent the inference of emotional attributes while maintaining the accuracy and utility of speech features. PSEGAN benefits from three modules: (1) A pre-trained speaker matcher imposes generative constraints on the model during the training phase to ensure that the generated speech retains the essential information needed for speaker recognition. (2) Attribute adversarial networks can generate perturbed speech that transforms emotional attributes while preserving the utility of the speech. (3) Gated Recurrent Networks (GRN) can handle the long-short term dependencies of speech signals. PSEGAN model solves the problem of utility loss in traditional speech privacy preservation methods based on generative adversarial networks (GAN). Experimental results show that on the RAVDESS dataset, PSEGAN reduces emotion recognition accuracy by 80.7%, while speaker recognition accuracy only decreases by 1.1%. These findings demonstrate that PSEGAN effectively mitigates the leakage of emotional attributes while maintaining high utility.

Soft-biometric privacy-enhancement methods are widely used in face recognition systems to prevent attackers from inferring soft-biometric attributes (e.g., gender, age, and race). However, existing methods typically focus on either representation-level or image-level privacy protection. In this paper, we propose a novel Face Conditional Privacy Funnel Network (Face-CPFNet), a dual-level privacy-enhancement framework with three key innovations. First, it introduces a dual-level privacy protection system where attackers receive reconstructed face images by combining the random guess of the sensitive attribute's potential label with the face representation provided by the user. Second, we propose the Face-CPF optimization problem, based on the Conditional Privacy Funnel (CPF) and an additional information leakage constraint. A deep variational approximation approach with parameterized deep neural networks is used to solve this problem and develop the Face-CPFNet model. Third, it enables supervised disentangled representation learning for reconstructing face images with variations in generative factors, and further enhances privacy protection by introducing an information leakage constraint when the sensitive attribute is a discrete binary random variable. Experimental results on benchmark datasets indicate that Face-CPFNet strikes a more effective balance between face verification accuracy and soft-biometric privacy compared with existing SBPE and CPF models.

Deep Neural Networks (DNNs) have been extensively employed for automatic face recognition, enabling the extraction of compact and discriminative representations from facial images. However, these representations typically encode a multitude of information ranging from individual identities to sensitive soft-biometric attributes such as gender, race, or age. This raises concerns regarding the privacy disclosure of soft-biometric as these attributes should be protected. To address this issue, we propose a novel Face Information Bottleneck Network (FIBNet), which is a representation-level privacy-enhancing framework based on the Information Bottleneck (IB) principle. The proposed FIBNet differs significantly from previous representation-level privacy-enhancing techniques in three key aspects. First, it generates a privacy-enhanced face representation, providing novel insights through an information-theoretic privacy framework. Second, we formulate the privacy protection of soft-biometric attributes as an IB optimization problem by striking a tradeoff between preserving a controlled amount of identity information within face representations and suppressing soft-biometric attribute information. Last, the proposed approach protects soft-biometric privacy from adversaries interested in specific sensitive attributes that are unknown to the biometric system designers or users. Detailed experimental results obtained on widely recognized facial recognition datasets demonstrate that the proposed FIBNet significantly outperforms the state-of-the-art methods in terms of both biometric performance for face verification and its soft-biometric attribute suppression efficiency. These notable results verify FIBNet as a novel and effective approach for ensuring representation-level soft-biometric privacy.

Evaluating startup companies is an important management process for technology business incubators and it is also a typical multicriteria decision-making (MCDM) problem. There exist various methods that have proposed to solve MCDM problems, but these methods heavily depend on the exact criteria weight values. The decision results of these methods are unstable. Moreover, they cannot provide the improvement suggestions for the nonoptimal startup companies. To overcome these two drawbacks, we propose a novel hesitant fuzzy linguistic decision-making method to solve the problem of evaluating startup companies. To this end, a novel semantic comparison method based on the experts' psychology and the ratio of score value to deviation degree is proposed to compare the hesitant fuzzy linguistic term sets. Then, a novel definition of hesitant fuzzy linguistic information envelopment efficiency (HFLIEE) is proposed, based on which, a novel hesitant fuzzy linguistic information envelopment analysis (HFLIEA) model and a novel preference model are proposed. By solving these models, all the alternatives can be ranked and nonoptimal alternatives can be improved. Finally, the numerical analysis is given to illustrate the applicability of the proposed models and the robustness analyses of the proposed models are provided. At the same time, they are compared with the previous hesitant fuzzy linguistic decision-making methods.