With the popularization of face recognition technology in IoT-Cloud, the problem of privacy leakage caused by it is becoming more and more serious. Although traditional privacy protection schemes can improve security to a certain extent, there is still a risk of data leakage when facing semi-trusted cloud servers. To this end, this paper proposes an anonymized face verification scheme for IoT convergence scenarios, which achieves real-time retrieval and secure matching of dense face features in virtual device copies by combining homomorphic encryption (CKKS) and privacy information retrieval (PIR) for anonymized face verification. The scheme ensures that the semi-trusted cloud server cannot obtain user-specific index information and matching results. Experiments show that the scheme’s verification accuracy in the ciphertext state on the LFW dataset is consistent with the plaintext, up to 97.06%, and can complete a privacy-protected anonymized facial verification process within seconds. The scheme is feasible in security demanding scenarios.

Federated Learning (FL) has been widely applied across various domains for collaborative training while preserving data privacy, but it remains highly vulnerable to poisoning attacks that compromise the global model’s integrity and performance. Existing clustering-based defense methods, such as FLAME, filter out malicious models by calculating vector similarities between local models but rely heavily on the assumption of independent and identically distributed client data. In this paper, we propose KSFed, a novel defense framework that detects poisoning attacks through the analysis of probability distributions of model parameters. KSFed treats model parameters as samples from a distribution, where benign models exhibit similar patterns while malicious models show significant deviations, allowing it to identify and filter out malicious models without relying on assumptions about attack types, adversarial strategies, or client data distributions. Experimental results show that KSFed surpasses FLAME and other clustering-based defenses, reducing backdoor accuracy (BA) by 8.33% to 99.74% under sophisticated attack strategies and highly non-IID client data distributions, while maintaining the global model’s main task accuracy (MA).

Artificial intelligence (AI) on edge refers to combining edge computing and AI, and enjoys the benefit of distributed structure, intelligence, and timeliness. Specifically, natural language processing model, which allows to processing language data right close to the device location within milliseconds and providing intelligent controller, have revolutionized researches. Recently, privacy concerns spiked when it is applied in healthcare, autonomous vehicles, manufacturing, etc. Secure multi-party computation has the advantage of strong security guarantee and computability over multi-sourced data. However, it is challenging to translate the timeliness benefit of Edge AI to secure deployment, as only constrained computation and storage resources are available for the edge nodes. We focus on the natural language processing (NLP), and design an efficient secure three-party computation protocol (called QuickNLP) in the semi-honest setting tolerating one corruption. Specifically, for the non-linear operations, we adopt the constant-round distributed comparison function (${\sf DCF}$) to evaluate the piecewise function efficiently with high accuracy. The proposed framework has been experimented with Python and the results show that QuickNLP could be a valuable solution for data privacy in edge computing. Specifically, compared to the existing protocol, we improve the computation costs by a factor of roughly$7\times$.

Linear code, a foundational construct extensively employed in communication, data transmission, and error correction, has been the subject of rigorous study for decades. Despite significant academic successes and widespread adoption, recent studies show that decoding methods for general linear codes, such as syndrome decoding, require the storage and search of large decoding tables, leading to inefficiencies. To mitigate this, Debris-Alazard et al. first proposed adapting Babai's algorithm and the LLL algorithm from lattice theory to binary codes, achieving considerable performance gains. Inspired by this, in this paper, we aim to explore the design of more general linear codes to overcome the limitations of baseline binary codes and enhance their applicability in more advanced applications such as DNA storage and 5G communication systems. To address this gap, we extend the foundational domain and decoding algorithms from lattices to q-ary codes. Specifically, we define a new fundamental domain and propose a polynomial-time decoding algorithm, RedtoFun. To validate our findings, we conduct a series of experiments to evaluate its real-world performance. The results demonstrate that our optimized RedtoFun algorithm surpasses the syndrome decoding scheme in terms of memory overhead and runtime while maintaining performance on par with the SizeRed decoding scheme.

Secure messaging heavily relies on a session key negotiated by an Authenticated Key Exchange (AKE) protocol. However, existing AKE protocols only verify the existence of a random secret key (corresponding to a certificated public key) stored in the terminal, rather than a legal user who uses the messaging application. In this paper, we propose a Biometrics-Authenticated Key Exchange (BAKE) framework, in which a secret key is derived from a user's biometric characteristics that are not necessary to be stored. To protect the privacy of users' biometric characteristics and realize one-round key exchange, we present an Asymmetric Fuzzy Encapsulation Mechanism (AFEM) to encapsulate messages with a public key derived from a biometric secret key, such that only a similar secret key can decapsulate them. To manifest the practicality, we present two AFEM constructions for two types of biometric secret keys and instantiate them with irises and fingerprints, respectively. We perform security analysis of BAKE and show its performance through extensive experiments.