Опубликована: Май 9, 2024
Язык: Английский
Computer Communications, Год журнала: 2024, Номер 222, С. 161 - 176
Опубликована: Апрель 24, 2024
Recently, big data related to human movement, air quality, and meteorology have been generated in urban computing through sensing technology the infrastructure. However, security problems arise as utilization increases. If from internet of things devices are constantly exposed, users' private information can be determined, a critical risk that could result privacy breaches. This paper proposes secure processing system using blockchain differential for protection computing. When service provider requests information, generates it machine learning. We apply these protect privacy. if query repeats, may provide insufficient protection. Therefore, we reduce total cost by reusing noise same parameters blockchain. Machine learning accuracy decrease when noisy used training. Thus, increase storing appropriately model design, simulate, analyze results an experimental environment parameter The proposed approach reduces costs compared existing mechanism while protecting demonstrate that, utilization, improves conventional mechanisms.
Язык: Английский
Процитировано
3
Mathematics, Год журнала: 2024, Номер 12(13), С. 1993 - 1993
Опубликована: Июнь 27, 2024
Although federated learning is gaining prevalence in smart sensor networks, substantial risks to data privacy and security persist. An improper application of techniques can lead critical breaches. Practical effective privacy-enhanced (PEPFL) a widely used framework characterized by low communication overhead efficient encryption decryption processes. Initially, our analysis scrutinized vulnerabilities within the PEPFL identified an attack strategy. This strategy enables server derive private keys from content uploaded participants, achieving 100% success rate extracting participants’ information. Moreover, when number participants does not exceed 300, time surpass 3.72 s. Secondly, this paper proposes model that integrates homomorphic secret sharing. By using sharing among instead secure multi-party computation, amount information available servers reduced, thereby effectively preventing inferring gradients. Finally, scheme was validated through experiments, it found significantly reduce inherent collusion unique scenario. even if some are unavailable, reconstructable nature ensures process continue uninterrupted, allowing remaining users proceed with further training. Importantly, proposed exerts negligible impact on accuracy
Язык: Английский
Процитировано
3
Peer-to-Peer Networking and Applications, Год журнала: 2024, Номер unknown
Опубликована: Сен. 7, 2024
Язык: Английский
Процитировано
3
Internet of Things, Год журнала: 2023, Номер 24, С. 100956 - 100956
Опубликована: Сен. 28, 2023
Decentralized Federated Learning (DFL) is a prevalent approach to efficiently train deep learning models and preserve privacy by sharing model gradients instead of the local data. However, participants in DFL may opt adopt dynamic behavior for personal gains. The existing cannot differentiate between adaptive massively distributed environments assume that all are honest. As result, free riders or malicious remain undetected not penalized. In this paper, we present architecture where decentralized assess each other using quality gradients. A novel reputation assessment protocol implemented detect eliminate with behavior. proposed evaluated behavior-based attacks environment increasing percentage from 10% 40%. results show our can effectively behaviour only two rounds whereas centralized federated fails attacks.
Язык: Английский
Процитировано
9
Опубликована: Май 9, 2024
Язык: Английский
Процитировано
2