2026· IEEE Transactions on Information Forensics and Security· Vol 21, pp. 7287-7300· 0 citations· 34 references
Abstract
Secure Federated Learning protects user privacy by keeping local gradients within the ciphertext space, thereby preventing direct exposure of raw gradients during aggregation. Paradoxically, this very protection introduces a fundamental vulnerability: the opacity of encrypted gradients blinds the server, rendering honest and malicious gradients indistinguishable. An attacker can exploit this opacity to inject manipulated gradients directly into the secure aggregation process, compromising the integrity of the global model. To address this issue, we propose SFLBR, a novel Byzantine-robust Secure Federated Learning framework. Specifically, we adopt the gradient median as a robust benchmark and introduce a layer-wise cosine similarity metric to differentiate malicious gradients. Furthermore, considering the complexity of non-independent and identically distributed (Non-IID) scenarios, we further design a proactive defense strategy that constrains the divergence among honest gradients, and augment it with a trust score function to improve robustness against potential misjudgments. Finally, we construct a series of secure communication protocols based on a lightweight dual-masking encryption mechanism, which enable efficient robust aggregation directly within the ciphertext space. Extensive experiments validate the effectiveness of SFLBR, demonstrating state-of-the-art (SOTA) defense performance across diverse attack scenarios.
Federated Learning (FL) enables collaborative model training without centralizing client data, making it well-suited for privacy-sensitive domains. Existing approaches use techniques such as homomorphic encryption, differential privacy, and secure multi-party computation to mitigate attacks including model inversion, membership inference, and gradient leakage. However, these methods often incur high computational and memory overheads and frequently overlook confidentiality of the global model itself, which may be proprietary and sensitive. These limitations reduce the practicality of secure FL in large-scale and compliance-sensitive environments.We present NETFL, a Fully Secure and scalable FL framework that decentralizes training across client pairs using lightweight MPC, while restricting servers to secure aggregation, client pairing, and routing. NETFL eliminates server-side bottlenecks, avoids full data offloading, and preserves confidentiality of data, model parameters, and updates throughout training. Our evaluation shows that NETFL protects against unauthorized observation, reconstruction, gradient leakage, membership inference, and inversion attacks, while achieving up to 13× faster training time and 50% lower server memory usage compared to prior work.
Sahar Ghoflsaz Ghinani, Elaheh Sadredini· International Conference on...· 0 citations
This work proposes PRoVeFL-a novel, modular FL framework that is Privacy-preserving, Byzantine-Robust, and ensures Verifiable aggregation, and improves runtime over the prior works, Prio and ELSA, based on distributed trust with comparable security guarantees, up to 100x and 10x, respectively.
Harsh Kasyap, Anil Kumar Pradhan, U. Atmaca et al.· 0 citations
Deep learning is becoming popular in cloud applications and serves to provide intelligent services; data aggregation in a central location makes sensitive information vulnerable to privacy breaches, regulatory infractions, and adversarial manipulation. All modern privacy mechanisms offer partial protection and frequently lack accuracy, scalability, or practicality in their operations. To overcome these limitations, a federated deep learning model is formulated so that secure joint learning can occur without transferring raw data across the domains of ownership. The framework incorporates training that is decentralized, training that uses differential privacy, training that uses secure aggregation, training that uses encrypted communication, and training that uses trust-based anomaly defense to defend against leakage, poisoning, and inference attacks. It also supports heterogeneous and highly non-IID datasets using adaptive coordination and stability-relevant participation regulation and meets emerging data protection requirements. The methods of resource-conscious orchestration and the optimization of communication eliminate overhead without obstructing the effectiveness of learning. The paradigm has therefore formed a privacy-by-design intelligent cloud ecosystem which ensures confidentiality, maintains performance, enhances robustness, and ensures responsible AI implementation in privacy-related sectors of healthcare, finance, governance, and smart infrastructure.
Sribidhya Mohanty, Pallavi Gupta, Anil Pratap Singh et al.· 2026 International Conferenc...· 0 citations
This work proposes a lightweight and information-theoretically secure aggregation framework that securely computes the majority vote (MV) polynomial through single-round secure multiplication, ensuring end-to-end information-theoretic security under the honest-majority assumption while revealing only the final aggregated sign to the server.
Federated learning alleviates data silos through a “data-local, model-global” paradigm, but transmitting plaintext gradients exposes clients to reconstruction attacks from malicious servers. Existing secure aggregation methods face trade-offs among privacy, accuracy, and efficiency: homomorphic encryption incurs high overhead, differential privacy sacrifices accuracy, and lightweight secret-sharing schemes often lack weighted aggregation support and suffer accuracy degradation as client numbers grow. To address these limitations, we propose SecAGG, a lossless secure weighted aggregation scheme based on additive secret sharing. SecAGG adopts a three-tier architecture consisting of client clusters, cooperative servers, and a super server. Clients split weighted model parameters into random shares and distribute them to cooperative servers, which perform encrypted partial aggregation before the super server securely reconstructs the global model. Experimental results demonstrate that SecAGG achieves strict security against up to M-1 colluding servers under the semi-honest model while preserving FedAvg-equivalent accuracy with minimal computation and communication overhead, effectively balancing privacy, accuracy, and efficiency.
Xiaomei Tian· 2026 3rd World Conference on...· 0 citations
Federated learning (FL) has emerged as an important distributed machine learning paradigm allowing many users to train a model together without sharing raw data. However, FL’s decentralized design makes it very susceptible to both Byzantine and coordinated model poisoning attacks, where a few malicious rounds of gradients are strategically inserted to reduce the overall integrity of the model. Even the existing Byzantine-resilient aggregation methods such as geometric median, Krum, and trimmed mean are vulnerable to sophisticated, coordinated poisoning attacks that leverage statistical gaps in outlier detection. In this article, we present FedSentinel, a novel Byzantine-resilient federated learning framework that combines cryptographic gradient attestation with adaptive trust-weighted aggregation to protect against coordinated model-poisoning attacks, which are among the most serious challenges. Three key innovations are introduced in FedSentinel Cryptographic Gradient Attestation Protocol (CGAP) that utilizes commitment schemes and zero-knowledge range proofs to ensure the integrity of the gradients and verify that submitted updates satisfy predefined norm and direction constraints; in the current protocol configuration, CGAP provides gradient integrity verification rather than full gradient confidentiality from the server Dynamic Trust-Weighted Robust Aggregation (DT-RoA): the computation and updating of per-client trust scores based on historical gradient consistency and cross-validation signals; and Coordinated Attack Detection Engine (CADE): based on spectral analysis of the gradient covariance matrices, it detects and isolates colluding Byzantine clients. By participating in up to 100 clients on CIFAR-10, CIFAR-100, FEMNIST, and Sentiment140 datasets, extensive experiments conducted under up to 30% Byzantine adversaries reveal that FedSentinel achieves 91.36% average global accuracy, whereas state-of-the-art defenses such as FLTrust, FLAME, RoFL, ShieldFL, and DnC achieve 83.83–86.94%. FedSentinel outperforms these defenses by 4.42–7.53% in terms of accuracy, while decreasing attack success rates by 53.2% under coordinated Byzantine backdoor attacks. The proposed framework offers a promising approach for federated learning that is verifiable and trustworthy in adversarial environments.
Abdullah Abdulkarim Alnajim· Electronics· 0 citations