Search

Federated learning (FL) enables multiple data sources to collaboratively train a global model for Multi-source Visual Fusion and Understanding (MSVFU) without centralizing raw data. However, it is difficult for such a global model to perform optimally due to the label distribution skew of images from various data sources. This paper presents FedCPD, a novel federated learning framework that effectively mitigates label distribution skew, enhancing global model accuracy and generalization capabilities through Class Proxy Decoupling aggregation and Proxy Regularization. First, Class Proxy Decoupling aggregation effectively decouples observed class proxies from missing ones in client updates, ensuring that the global model's aggregation process is not adversely affected by inaccurate class proxy updates. Second, Proxy Regularization proactively increases the attention of class proxies to features of other classes during local training, thereby enhancing the model's generalization capability across diverse data sources. Additionally, we integrate a pre-trained Vision Transformer (ViT) feature extractor to enhance the global model's robustness against label distribution skew. Extensive evaluation on four public datasets with varying label distribution skew confirms the superior efficacy of our approach compared to existing methods. • Class proxy decoupling aggregation tackles label skew in federated learning. • Proxy regularization enhances cross-class learning in federated learning. • Integration of advanced feature extractor strengthens model against data skew. [ABSTRACT FROM AUTHOR]

Social network data can help with obtaining valuable insight into social behaviors and revealing the underlying benefits. New big data technologies are emerging to make it easier to discover meaningful social information from market analysis to counterterrorism. Unfortunately, both diverse social datasets and big data technologies raise stringent privacy concerns. Adversaries can launch inference attacks to predict sensitive latent information, which is unwilling to be published by social users. Therefore, there is a tradeoff between data benefits and privacy concerns. In this paper, we investigate how to optimize the tradeoff between latent-data privacy and customized data utility. We propose a data sanitization strategy that does not greatly reduce the benefits brought by social network data, while sensitive latent information can still be protected. Even considering powerful adversaries with optimal inference attacks, the proposed data sanitization strategy can still preserve both data benefits and social structure, while guaranteeing optimal latent-data privacy. To the best of our knowledge, this is the first work that preserves both data benefits and social structure simultaneously and combats against powerful adversaries. [ABSTRACT FROM PUBLISHER]

The huge amount of sensory data collected from mobile devices has offered great potentials to promote more significant services based on user data extracted from sensor readings. However, releasing user data could also seriously threaten user privacy. It is possible to directly collect sensitive information from released user data without user permissions. Furthermore, third party users can also infer sensitive information contained in released data in a latent manner by utilizing data mining techniques. In this paper, we formally define these two types of threats as inherent data privacy and latent data privacy and construct a data-sanitization strategy that can optimize the tradeoff between data utility and customized two types of privacy. The key novel idea lies that the developed strategy can combat against powerful third party users with broad knowledge about users and launching optimal inference attacks. We show that our strategy does not reduce the benefit brought by user data much, while sensitive information can still be protected. To the best of our knowledge, this is the first work that preserves both inherent data privacy and latent data privacy. [ABSTRACT FROM AUTHOR]

The rising popularity of mobile social media enables personalization of various content sharing and subscribing services. These two types of services entail serious privacy concerns not only to the confidentiality of shared content, but also to the privacy of end users such as their identities, interests and social relationships. Previous works established on the attribute-based encryption (ABE) can provide fine-grained access control of content. However, practical privacy-preserving content sharing in mobile social networks either incurs great risk of information leaking to unauthorized third parties or suffers from high energy consumption for decrypting privacy-preserving content. Motivated by these issues, this paper proposes a publish-subscribe system with secure proxy decryption (PSSPD) in mobile social networks. First, an effective self-contained privacy-preserving access control method is introduced to protect the confidentiality of the content and the credentials of users. This method is based on ciphertext-policy ABE and public-key encryption with keyword search. After that, a secure proxy decryption mechanism is proposed to reduce the heavy burdens of energy consumption on performing ciphertext decryption at end users. The experimental results demonstrate the efficiency and privacy preservation effectiveness of PSSPD. [ABSTRACT FROM AUTHOR]

We consider the problem of tracking quantiles and range countings in wireless sensor networks. The quantiles and range countings are two important aggregations to characterize a data distribution. Let S ( t ) = ( d 1 , … , d n ) denote the multi-set of sensory data that have arrived until time t , which is a sequence of data orderly collected by nodes s 1 , s 2 , … , s k . One of our goals is to continuously track ϵ -approximate ϕ -quantiles ( 0 ≤ ϕ ≤ 1 ) of S ( t ) for all ϕ 's with efficient total communication cost and balanced individual communication cost. The other goal is to track ( ϵ , δ ) -approximate range countings satisfying the requirement of arbitrary precision specified by different users. In this paper, a deterministic tracking algorithm based on a dynamic binary tree is proposed to track ϵ -approximate ϕ -quantiles, whose total communication cost is O ( k / ϵ ⋅ log ⁡ n ⋅ log 2 ⁡ ( 1 / ϵ ) ) , where k is the number of the nodes in a network, n is the total number of the data, and ϵ is the user-specified approximation error. For range countings, a Bernoulli sampling based algorithm is proposed to track ( ϵ , δ ) -approximate range countings, whose total communication cost is O ( 2 ϵ 2 ln ⁡ 2 1 − 1 − δ + n c ) , where δ is the user-specified error probability, n c is the number of clusters. [ABSTRACT FROM AUTHOR]

User representation learning is one prominent and critical task of user analysis on social networks, which derives conceptual user representations to improve the inference of user intentions and behaviors. Previous efforts have shown its substantial value in multifarious real-world applications, including product recommendation, textual content modeling, link prediction, and many more. However, existing studies either underutilize multi-view information, or neglect the stringent entanglement among underlying factors that govern user intentions, thus deriving deteriorated representations. To overcome these shortages, this paper proposes an adversarial fusion framework to fully exploit substantial multi-view information for user representation, consisting of a generator and a discriminator. The generator learns representations with a variational autoencoder, and is forced by the adversarial fusion framework to pay specific attention to substantial informative signs, thus integrating multi-view information. Furthermore, the variational autoencoder used in the generator is novelly designed to capture and disentangle the latent factors behind user intentions. By fully utilizing multi-view information and achieving disentanglement, our model learns robust and interpretable user representations. Extensive experiments on both synthetic and real-world datasets demonstrate the superiority of our proposed model. • An adversarial fusion framework preserves substantial multi-view information. • A Disentangled Variational Autoencoder learns interpretable representations. • A holistic optimization strategy balances the training of different modules. • Our proposed model outperforms the state-of-the-art w.r.t opinion prediction task. [ABSTRACT FROM AUTHOR]

The social activities of Primary Users (PUs) and Secondary Users (SUs) affect actual accessible whitespace in Cognitive Radio Networks (CRNs). However, the impacts of primary activities on available whitespace have been extensively investigated due to the dominating priority of PUs, while the impacts of secondary activities on actual accessible whitespace have been ignored. Therefore, we propose to incorporate the primary and secondary activities in the analysis and decision of the accessible whitespace, namely, both the dominance of PUs over SUs and the competitions among SUs are simultaneously taken into account. Specifically, we first approximate primary activity probability based on the real datasets of mobile phone usage records, then the spectrum opportunity between a pair of communication SUs is deduced based on primary activities. Next, we infer the access probability limit of SUs successfully accessing the whitespace according to the primary activity probability, and depict the secondary activity probability from the views of social activity patterns and social networks respectively. Furthermore, the actual accessible probability of whitespace is given by introducing the competitions among SUs. Finally, a greedy routing algorithm, considering the accessible whitespace and the distance to the destination, is proposed to verify our idea. The experiment results based on the real datasets demonstrate the correctness of our analysis and the advantages of the proposed algorithm. [ABSTRACT FROM AUTHOR]